Systems and methods for automated assessment of artificial tissue

Abstract

Systems and methods are described herein for geometric encoding of tissue phenotype. At least in some aspects, the systems and methods comprise obtaining an image of an artificial tissue under a first set of conditions. The systems and methods further compromise extracting a first region from the image such that the first region circumscribes the artificial tissue within the image. The systems and methods further compromise extracting one or more geometric features from the first region, wherein the one or more geometric features provide a shape-based characterization of the artificial tissue. The system and methods further comprise generating a first tissue descriptor based on the one or more geometric features. The first tissue descriptor can encode a phenotype of the artificial tissue under the first set of conditions. Still further, in some aspects, the systems and method may comprise implementation of quality control (QC) of artificial tissue.

Description

[0001] SYSTEMS AND METHODS FOR AUTOMATED ASSESSMENT OF ARTIFICIAL TISSUE

[0002] RELATED APPLICATION

[0003] This application claims the benefit of U.S. Provisional Application No. 63 / 525,451 (filed on July 7, 2023), which is incorporated by reference herein.

[0004] BACKGROUND

[0005] Quality control (QC) of an artificial, or engineered, tissue can be used during maturation of the artificial tissue to track the development of the artificial tissue determine whether to include or exclude the artificial tissue from further analysis. Existing QC approaches typically rely on manual inspection of the artificial tissue during the maturation process. Such manual inspection leads to user-induced bias which can result in high variability in the quality of the artificial tissue generated. This can have a detrimental impact on any downstream tasks which rely on such artificial tissue, such as drug discovery and development tasks. Moreover, manual QC inspection can inhibit the large scale production of artificial tissue. In addition, even when artificial tissue passes QC inspection, it is often difficult to identify or explain variability arising in experimental outcomes when artificial tissue is used within experimental settings or downstream tasks.

[0006] Therefore, there is a need for new approaches to assessing the quality of artificial tissue (e.g., during maturation) and encoding the condition dependent phenotype of artificial tissue.

[0007] SUMMARY OF DISCLOSURE

[0008] According to an aspect of the present disclosure there is provided a method for geometric encoding of tissue phenotype. The method comprises obtaining an image of an artificial tissue under a first set of condition, extracting a first region from the image such that the first region circumscribes the artificial tissue within the image, and extracting one or more geometric features from the first region. The one or more geometric features provide a shape-based characterization of the artificial tissue. The method further comprises generating a first tissue descriptor based on the one or more geometric features, wherein the first tissue descriptor encodes a phenotype of the artificial tissue under the first set of conditions.

[0009] According to a further aspect of the present disclosure there is provided a system for quality control (QC) of artificial tissue. The system comprises a bioreactor comprising a device configured for growing tissue, and a sensor assembly configured to obtain one or more images of a tissue within the device. The system further comprises a QC unit communicatively coupled to the bioreactor. The QC unit comprising one or more processors configured to obtain, from the bioreactor at a first predetermined time point, a first image of the tissue within the device and extract one or more image segments from the first image, wherein each of the one or more image segments comprise a region of interest of the tissue within the first image. The one or more processors of the QC unit are further configured to determine one or more tissue parameters from the one or more image segments, wherein the one or more tissue parameters are indicative of a physiological state of the tissue at the first predetermined time point. The one or more processors of the QC unit are further configured to output a QC report based on the one or more tissue parameters.

[0010] According to an additional aspect of the present disclosure there is provided a system for engineered tissue attachment grading. The system comprises a multi-task learning (MTL) network comprising an input convolutional neural network configured to receive an input image, a plurality of independent output networks each configured to estimate an attachment score associated with the input image, and a shared network coupled between the input convolutional neural network and each of the plurality of independent output networks. The system further comprises a control unit communicatively coupled to the MTL network and comprising one or more processors configured to obtain an image of an engineered tissue grown within a device comprising a tissue scaffold for attachment to the engineered tissue and extract a first region from the image, wherein the first region of the image includes a first portion of the engineered tissue and a portion of the tissue scaffold. The one or more processors of the control unit are further configured to determine, using the MTL network, a plurality of attachment scores based on the first region and determine an attachment grading for the first region based on the plurality of attachment scores, wherein the attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to an attachment site of the device. The one or more processors of the control unit are further configured to output the attachment grading.

[0011] In accordance with the above, and with the disclosure herein, the present disclosure includes applying certain features or aspects with or by use of, a particular machine, e.g., a bioreactor. In various aspects, the bioreactor may comprise a device configured for growing or manipulating human tissue (e.g., human body tissue such as muscle tissue, cardiac tissue, and / or skeletal muscle tissue). Additionally, or alternatively, bioreactor may comprise a sensor assembly configured to assess the quality of artificial tissue (e.g., during maturation) and encoding the condition dependent phenotype of artificial tissue.

[0012] Further, the present disclosure includes effecting a transformation or reduction of a particular article to a different state or thing, e.g., the transformation or reduction of image(s) of human tissue, e.g., within a bioreactor as sensed by a sensor assembly, to a different state or thing, e.g., the generation, creation, or otherwise development of a QC report based on an originally received imaging data as sensed (e.g., by one or more sensors) from an artificial tissue, which is then transformed or reduced into segments and features as part of the imaging analysis.

[0013] Still further, the present disclosure includes improvements in computer functionality or in improvements to other technologies at least because the disclosure herein discloses systems and methods for reducing error in underlying computing devices, e.g., by determining the quality and flagging issues of engineered tissue within a bioreactor. The automated and image-based systems and methods herein reduce error, particularly during the maturation period of artificial tissue, for example, by decreasing user assessment bias. This improves the underlying system as it would have fewer false positives (when compared to conventional quality control observation) because it would rely on imagebased processing utilizing feature extraction to specifically identify artificial tissues or phenotypes thereof. This also allows for generation of more accurate and less error prone output, and, as a result, high-quality drug products, such as therapeutics.

[0014] In addition, the present disclosure relates to improvement to other technologies or technical fields at least because the systems and methods of the present disclosure provide or otherwise use a two-step segmentation process to perform segmentation of regions from an image of an artificial tissue. The two-step segmentation process provides an efficient and accurate approach for identifying the tissue regions This in turn leads to improvements in the efficiency, accuracy, and overall performance over the underlying system. Particularly, the machine learning models used in an extraction unit can be more efficiently trained in less time, and to have more predictive accuracy, when the tissue regions are more accurately identified. Further, because of this, the accuracy of the machine learning model is improved, which improves the efficiency and performance of downstream drug discovery and development tasks.

[0015] Still further, the systems and methods, when deployed on the underlying computing device, allows the systems and methods of the present disclosure to execute with fewer iterations, and use fewer computing resources, than prior art related systems and methods. That is, the present disclosure describes improvements in the functioning of the computer itself or "any other technology or technical field" because the underlying computing device can generate a tissue descriptor based on geometric feature(s) where, for example, a given tissue descriptor can encode a phenotype of the artificial tissue under. The encoding of tissue phenotype provided by the tissue descriptor provides a powerful and discriminative feature which can be efficiently used for a number of downstream drug discovery and development tasks, including reduction of a number of test cycles and resources expended therefrom. That is the improvement provided by the generation of the tissue descriptor based on geometric feature(s) allows the underlying computer system to utilize less processing and memory resources compared to prior art systems and methods. This at least because such features can be used to generate or determine a phenotype of an artificial tissue under a set of conditions, without the need for various tests and / or empirical computer simulation across a wide range of tests using multiple compute cycles and data. Said another way, the systems and methods of the present disclosure improve over the prior art at least because prior art systems and methods require an empirical or trial-and-error approach that can involve real-world trials on human tissue that can result in, and require, large database and memory utilization and processor usage to arrive at a similar real-world or simulated results that has a same or similar result. On the other hand, the disclosed systems and methods describe generation and / or use of a bioreactor for growing and testing tissue for defining a limited set of data specific to the tissue (e.g., human engineered tissue), which requires less memory usage and / or processing utilization compared to a conventional approach where large sets of unknown, potentially irrelevant data is used or required. Moreover, the disclosure herein allows for identification and use of high-fidelity data sets, which reduces the need for additional computational cycles further. Moreover, the tissue descriptors can supplement existing tissue assays to improve the explainability of such models in the presence of experimental variability (e.g., by providing quantitative explanations of the physiological or morphological state of the artificial tissue).

[0016] Still further, the present disclosure includes specific features other than what is well- understood, routine, conventional activity in the field, and / or otherwise adds unconventional steps that confine the disclosure to a particular useful application, e.g., systems and methods for automated assessment of artificial tissue, which can be used, for example, to automate assessment of artificial or engineered tissue quality (quality control) based on one or more tissue parameters, and / or preform automated assessment of artificial or engineered tissue phenotype based on one or more geometric tissue features, which can improve downstream tasks such as drug discovery and development. Advantages will become more apparent to those of ordinary skill in the art from the following description of the preferred embodiments which have been shown and described by way of illustration. As will be realized, the present embodiments may be capable of other and different embodiments, and their details are capable of modification in various respects. Accordingly, the drawings and description are to be regarded as illustrative in nature and not as restrictive.

[0017] BRIEF DESCRIPTION OF DRAWINGS

[0018] Embodiments of the present disclosure will now be described, by way of example only, and with reference to the accompanying drawings, in which :

[0019] Figure 1 shows a system for automated assessment of an artificial tissue according to an aspect of the present disclosure;

[0020] Figure 2 shows an image of an artificial tissue according to an embodiment of the present disclosure;

[0021] Figure 3 illustrates a first segmentation step of a morphological based segmentation pipeline performed by a segmentation unit of a system for automated assessment of an artificial tissue according to an embodiment of the present disclosure;

[0022] Figure 4 illustrates a second segmentation step of a morphological based segmentation pipeline performed by a segmentation unit of a system for automated assessment of an artificial tissue according to an embodiment of the present disclosure;

[0023] Figure 5 illustrates features which are extracted from tissue regions of a segmented image according to an aspect of the present disclosure;

[0024] Figure 6 shows the decomposition of a well-cropped image of an artificial tissue for tissue attachment grading according to an embodiment of the present disclosure;

[0025] Figure 7 shows a hierarchical tissue attachment grading according to embodiments of the present disclosure;

[0026] Figure 8 shows cropped regions of three artificial tissues having different predicted tissue attachment gradings according to embodiments of the present disclosure;

[0027] Figures 9A-9B show a multi-task learning (MTL) network for tissue attachment grading according to an aspect of the present disclosure; Figure 10 shows a sequence of images of an artificial tissue during a maturation period according to an embodiment of the present disclosure;

[0028] Figure 11 shows a portion of a QC report obtained over the maturation period of the artificial tissue shown in Figure 10 according to an embodiment of the present disclosure;

[0029] Figure 12 shows a tissue attachment grading portion of a QC report obtained over the maturation period of the artificial tissue shown in Figure 10 according to an embodiment of the present disclosure;

[0030] Figure 13 illustrates a change in tissue phenotype as encoded by a geometric tissue descriptor according to an embodiment of the present disclosure;

[0031] Figure 14 shows a method for quality control (QC) of artificial tissue according to an aspect of the present disclosure;

[0032] Figure 15 shows a method for engineered tissue attachment grading according to an aspect of the present disclosure;

[0033] Figure 16 shows a method for geometric encoding of tissue phenotype according to an aspect of the present disclosure;

[0034] Figure 17 shows a method for determining an effect associated with a phenotypic change according to an embodiment of the present disclosure;

[0035] Figure 18 shows a bioreactor system according to embodiments of the present disclosure; and

[0036] Figure 19 shows an example computing system according to embodiments of the present disclosure.

[0037] TECHNICAL FIELD

[0038] The present disclosure relates to artificial or engineered tissue. Particularly, but not exclusively, the present disclosure relates to the automated assessment of artificial or engineered tissue quality based on one or more tissue parameters. Additionally, but not exclusively, the present disclosure relates to the automated assessment of artificial or engineered tissue phenotype based on one or more geometric tissue features. DETAILED DESCRIPTION

[0039] Figure 1 shows a system 100 for automated assessment of an artificial tissue according to an aspect of the present disclosure.

[0040] The system 100 comprises a bioreactor 102, a segmentation unit 104, an extraction unit 106, a quality control (QC) unit 108, and an assay unit 110 (alternatively referred to as an analysis unit, evaluation unit, or assessment unit). In various aspects, the bioreactor is used for growing tissue (e.g., engineered tissue), such as human body tissue including, by way of non-limiting example, muscle tissue, cardiac tissue, skeletal muscle tissue, or other human tissue). The bioreactor 102 comprises an artificial tissue 112 under a first set of conditions and a sensor assembly 114 configured to obtain an image 116 of the artificial tissue 112 at a first time point. The segmentation unit 104 extracts one or more regions (or segments) from the image 116 of the artificial tissue 112. Features or parameters associated with the artificial tissue 112 are extracted from the one or more regions by the extraction unit 106. The QC unit 108 uses the features or parameters extracted by the extraction unit 106 to generate a QC report 118 which comprises information indicative of a state of the artificial tissue 112 (e.g., a physiological or morphological state) under the first set of conditions at the first time point. The assay unit 110 uses the features or parameters extracted by the extraction unit 106 to generate a tissue descriptor 120 which encodes a phenotype of the artificial tissue 112 under the first set of conditions at the first time point. Figure 1 further shows a first subsystem 122 and a second subsystem 124. The first subsystem 122 comprises the segmentation unit 104 and the extraction unit 106. The second subsystem 124 comprises the segmentation unit 104, the extraction unit 106, the QC unit 108, and the assay unit 110.

[0041] The system 100 shown in Figure 1 comprises an automated system for qualitative and quantitative assessment of artificial tissue(s). Beneficially, the system 100 can quantify the quality of artificial tissue(s) and flag artificial tissue(s) which do not pass quality control conditions. The automation of quality control, particularly during the maturation period of artificial tissue, helps to decrease the requirement for human involvement thereby decreasing user assessment bias and increasing the scalability of the system 100. Moreover, the geometric tissue descriptors utilized by the system 100 provide an efficient and discriminative approach for encoding tissue phenotype. The geometric tissue descriptors can thus be used to support a number of downstream tasks such as drug discovery and development. Moreover, the tissue descriptors can supplement existing tissue assays to improve the explainability of such models in the presence of experimental variability (e.g., by providing quantitative explanations of the physiological or morphological state of the artificial tissue). Each of the units shown in Figure 1 (e.g., the segmentation unit 104, the extraction unit 106, etc.) can be implemented using one or more computing systems such as that shown in Figure 19 (other computing systems can be used as well). Computing tasks discussed herein as being performed at and / or by one or more functional unit(s) (e.g., as described in relation to Figure 1) can instead be performed remote from the respective system, or vice versa. Such configurations can be implemented without deviating from the scope of the present disclosure. The use of computer-based systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. Computer-implemented operations can be performed on a single component or across multiple components. Computer-implemented tasks and / or operations can be performed sequentially or in parallel. Data and instructions can be stored in a single memory device or across multiple memory devices.

[0042] In one embodiment, the segmentation unit 104 and the extraction unit 106 form part of the first subsystem 122 which is separate from the bioreactor 102, the QC unit 108, and the assay unit 110. The first subsystem 122 comprises hardware and / or processing logic configured to receive or otherwise obtain data from the bioreactor 102 (e.g., the image 116) and output data to the QC unit 108 and / or the assay unit 110. The first subsystem 122 may be physically collocated with the bioreactor 102, the QC unit 108, and / or the assay unit 110, or the first subsystem 122 may be remotely located (e.g., the first subsystem 122 is deployed as a cloud service and is accessible via one or more application programming interfaces (APIs)).

[0043] In a further embodiment, the segmentation unit 104, the extraction unit 106, the QC unit 108, and the assay unit 110 form part of the second subsystem 124 which is separate from the bioreactor 102. The second subsystem 124 is communicatively coupled to the bioreactor 102 and comprises hardware and / or processing logic configured to receive or otherwise obtain data from the bioreactor 102 (e.g., the image 116). The second subsystem 124 may be physically collocated with the bioreactor 102 or the second subsystem 124 may be remotely located (e.g., the second subsystem 124 is deployed as a cloud service and is accessible via one or more application programming interfaces (APIs)).

[0044] As will be described in more detail below in relation to Figure 18, the bioreactor 102 comprises one or more devices for growing and / or sustaining artificial, or engineered, tissue under one or more conditions (e.g., the artificial tissue 112 under the first set of conditions). The artificial tissue 112 comprises engineered muscle tissue. In one embodiment, the engineered muscle tissue is engineered cardiac tissue; more particularly, the engineered cardiac tissue is a 3-dimensional (3D) construct of human induced pluripotent stem cell derived cardiomyocytes (CM) (iCells, Fujifilm / CDI). In an alternative embodiment, the engineered muscle tissue is engineered skeletal muscle tissue. The artificial tissue 112 within the bioreactor 102 is associated with one or more conditions such as reference, or control, conditions and perturbation conditions. In general, reference conditions refer to conditions which provide a baseline comparison to the perturbation conditions. In embodiments, a reference condition is a condition associated with a control setup or environment. A reference condition may correspond to an artificial, or engineered, tissue in its default, natural, or unaltered state (i.e., without dosage of a drug or agent). Alternatively, a reference condition may correspond to vehicle treated artificial tissue. Perturbation conditions refer to conditions in which the artificial tissue has been perturbed in some way. Examples of perturbation conditions include the administration of a drug or compound (i.e., a perturbant), a disease state, a different cell line, a physical perturbance applied to the artificial tissue, or a change to the environment of the artificial tissue. In the case of perturbations involving a drug or compound, the conditions may alternatively be referred to as treatment conditions and are further associated with an effect related to the drug or compound such as a mechanism of action or a toxicity. Given the range of different perturbation conditions, an artificial tissue may be associated with more than one perturbation condition (e.g., a diseased engineered tissue having been treated with a specific compound).

[0045] The sensor assembly 114 is configured to obtain an image, such as the image 116, of the artificial tissue 112 at a first time point (e.g., 1 day post-seeding, 2 weeks post-seeding, etc.). As will be described in more detail below, the image 116 is segmented and analyzed to determine one or more features, or descriptors, related to the artificial tissue 112 under the one or more conditions at the first time point.

[0046] Figure 2 shows an image 200 of an artificial tissue according to an embodiment of the present disclosure.

[0047] The image 200 is a brightfield image of an artificial tissue within a device obtained from a bioreactor, such as the bioreactor 102 shown in Figure 1. Several regions have been annotated within the image 200 including a well region 202, a first tissue scaffold region 204, a second tissue scaffold region 206, and a tissue region 208. As will be discussed in more detail blow, these regions are automatically identified and utilized in downstream tasks performed by the system 100 of Figure 1 (e.g., quality control, estimation of tissue phenotype, etc.).

[0048] The well region 202 comprises a portion of the image 200 which circumscribes the well within which the artificial tissue is grown from cells seeded therein. The first tissue scaffold region 204 and the second tissue scaffold region 206 comprise portions of the image 200 which contain the tissue scaffold(s) of the device containing the artificial tissue. In the device captured within the image 200, there are two tissue scaffolds shown, one within each tissue scaffold region. Each tissue scaffold shown in the image 200 comprises a flexible polymer wire disposed across the well. The tissue scaffolds are configured to: (a) permit attachment of the artificial tissue formed therebetween, and (b) deform in response to the contractile force exerted thereon by the artificial tissue.

[0049] The tissue region 208 circumscribes the artificial tissue within the image 200 (i.e., the dark region positioned at the center of the image 200). As can be seen, the tissue region 208 lies within the well region 202 and overlaps both the first tissue scaffold region 204 and the second tissue scaffold region 206. The efficient and accurate automatic extraction of the tissue region 208 from the image 200 is a challenging task due to the large variability in the size and shape of tissues captured across images, as well as the extraneous structures within an image (such as those seen in the image 200).

[0050] Referring once again to Figure 1, the segmentation unit 104 is configured to extract one or more of the regions described above in relation to Figure 2. The segmentation unit 104 utilizes a two-step segmentation process to perform segmentation of regions from an image of an artificial tissue. The first segmentation step segments the well region (i.e., the well region 202 shown in Figure 2) from the image 116. The second segmentation step segments the tissue region (i.e., the tissue region 208 shown in Figure 2) from the image 116 cropped to the well region identified in the first segmentation step. The segmentation unit 104 is configured to perform further segmentation steps to extract the tissue scaffold regions (e.g., the first tissue scaffold region 204 and the second tissue scaffold region 206 shown in Figure 2).

[0051] Beneficially, utilizing a two-step segmentation process provides an efficient and accurate approach for identifying the tissue regions shown in Figure 2. Particularly, by utilizing the well region as a common reference frame, the tissue scaffold regions and the tissue region are more accurately identifiable. This in turn leads to improvements in the efficiency, accuracy, and overall performance of the extraction unit 106, the QC unit 108 and the assay unit 110 which depend on the segmentations generated by the segmentation unit 104. Particularly, the machine learning models used in the extraction unit 106 (as described below) can be more efficiently trained in less time when the tissue regions are more accurately identified.

[0052] Any suitable image segmentation approach can be used to segment the well region and subsequently segment the tissue region. Example approaches include clustering based segmentation (e.g., fuzz c-means clustering, Gaussian Mixture Modeling, etc.), neural network based segmentation approaches, and shape based segmentation. In one embodiment, the two-step segmentation process is implemented using a morphological based segmentation pipeline. Beneficially, this pipeline is unsupervised requiring no prior training and is moreover robust to noise. The pipeline is also able to segment the various regions from an input image quickly and efficiently with less processor and memory requirements than supervised machine learning models (e.g., trained neural networks).

[0053] Figure 3 illustrates the first segmentation step of a morphological based segmentation pipeline performed by a segmentation unit, such as the segmentation unit 104 of the system 100, according to an embodiment of the present disclosure.

[0054] Figure 3 shows an input image 302, an equalized image 304, a first binarized image 306, an eroded image 308, a first filled image 310, a first single object image 312, and a rough crop image 314. Figure 3 further shows a second binarized image 316, an edge image 318, a second filled image 320, a second single object image 322, an image object with center image 324, an adjusted image 326, and an output image 328.

[0055] The morphological segmentation pipeline illustrated in Figure 3 comprises a plurality of sequential morphological operations and other image processing operations. At a general level, the pipeline comprises two stages. At the first stage (from the input image 302 to the rough crop image 314), a rough crop of the well region is extracted from the input image 302 and extraneous structures are removed. By identifying the well region within the input image 302, the image can be rotated to align to the axes of the well region thereby improving the fine-grained segmentation of the well region during the second stage. The second stage (from the second binarized image 316 to the output image 328) performs a fine crop of the well region by utilizing edge-based features within the rough crop.

[0056] At the first stage, the input image 302 is provided as input to the segmentation pipeline. As described above, the image comprises a brightfield image of an engineered tissue and includes several extraneous structures as well as the regions of interest (e.g., the well region, tissue region, etc.).

[0057] A histogram equalization operation is then applied to the input image 302 and results in the equalized image 304. The histogram equalization operation adjusts the histogram of pixel intensities within the input image 302 such that the intensity values are more evenly distributed across the full range of possible intensities. In consequence, the global contrast of the equalized image 304 is improved as compared to that of the input image 302. Beneficially, this helps highlight detail within the input image 302 and also allows for consistent application of latter morphological operations within the pipeline (i.e., the parameter values for the operations do not need to be adapted on a per image basis).

[0058] A binarization operation is applied to the equalized image 304 resulting in the first binarized image 306. As is known, binarization is the process of converting a grayscale image to a binarized representation thereof. That is, all pixels within the grayscale image are replaced with either a "0" or "1" depending on the pixel intensity value and the thresholding approach used. Typically, if a pixel has an intensity value above a predetermined threshold, then it is assigned a binary value of "1" (i.e., the pixel is considered foreground), else it is assigned a binary value of "0" (i.e., the pixel is considered background). The Otsu method is used to determine the predetermined threshold for the equalized image 304. Alternatively, the predetermined threshold is set at a predetermined value (e.g., the median pixel intensity value), or using an optimization approach such as variational minimax optimization. In yet a further alternative, the foreground and background regions of the image are identified using a clustering approach such as fc-means clustering, Gaussian Mixture Models, or fuzzy fc-means clustering.

[0059] A first cycle of morphological operations are then applied to the first binarized image 306 to identify the approximate well region and remove noise. The morphological operations comprise an erosion of the first binarized image 306 (as shown by the eroded image 308) followed by a filling operation applied to the eroded image 308 (as shown by the first single object image 312).

[0060] Erosion is a morphological operation which applies a structuring element— a matrix defining a shape of the neighborhood around a pixel which will be used to process the pixel— to all pixels within an image to generate an output. The erosion operation sets the value of a pixel in the output image to be the minimum value of all pixels within the neighborhood of the pixel within the input image (where the neighborhood is defined by the structuring element). The erosion operation applied to the first binarized image 306 thus removes small objects and lines (e.g., due to noise) within the first binarized image 306 so that only larger structures are maintained within the eroded image 308. In one implementation, the erosion operation is performed using a cross-shaped structured element with a square connectivity of one.

[0061] The filling operation, applied to the eroded image 308, performs a hole filling (or region filling) algorithm to the eroded image 308 to produce the first filled image 310. The hole filling algorithm fills holes within "hollow" shapes in the eroded image 308. A hole is a contiguous background region of the eroded image 308 (i.e., all pixels within the region have a value of "0") surrounded by a connected border of foreground pixels (i.e., the background region is circumscribed by pixels having a value of "1"). The hole filling algorithm fills such background regions with foreground pixels.

[0062] After the filling operation has been applied, the first filled image 310 comprises multiple connected components— sets of foreground pixels (pixel value of "1") which are connected in some manner (e.g., pixels are connected if they share a common edge or corner). The largest of these connected components corresponds to a rough segmentation of the well, as seen in the first filled image 310. The largest connected component is thus identified from the first single object image 312 and used as a rough mask for the well region within the input image 302. As the well region may not be aligned to the axes of the image, the major axis of the largest connected component is determined (as illustrated by the line within the first single object image 312) and the angle between the major axis and the vertical axis of the input image 302 is used to calculate the rotation to be applied to the input image to align the well region with the image axes. Both the first filled image 310 and the input image 302 are rotated by the calculated angle and the input image 302 is cropped to the bounding box of the largest component within the (rotated) filled image. This rough crop of the well region is shown in the rough crop image 314.

[0063] At the second stage, the rough crop image 314 is binarized to produce the second binarized image 316. An edge detection operation is then applied to produce the edge image 318. The edge image 318 corresponds to a transformation of the second binarized image 316 with connected edges. The edge image 318 is generated using a pair of dilation operations. A first dilation operation is applied to the second binarized image 316 using a rectangular structured element extending in a first direction (e.g., horizontal) and a second dilation operation is applied to the output of the first dilation operation using a rectangular structured element extending in a second direction (e.g., horizontal). In one implementation, the first structured element is a rectangular structured element of shape (5 x 1) and the second structure element is a rectangular structured element of shape (1 x 5).

[0064] Using similar operations as those described above in relation to the first filled image 310 and the first single object image 312, the edge image 318 is filled to produce the second filled image 320 and the largest connected component extracted as shown in the second single object image 322. The object with center image 324 shows the center of the single connected component of the second single object image 322 overlaid the input image 302 (masked using the second single object image 322). The adjusted image 326 shows the input image 302 rotated according to the transformation described above in relation to the first single object image 312 and the rough crop image 314. The center of the single connected component shown in the object with center image 324 is overlaid the adjust image 326.

[0065] The output image 328 comprises the portion of the input image 302 which contains the well region aligned such that the vertical axis of the well region aligns to the vertical axis of the output image 328. As described in more detail below, the output image 328 is then used to extract the tissue scaffold regions (e.g., the first tissue scaffold region 204 and the second tissue scaffold region 206 shown in Figure 2). The output image 328 is also used as the input to the second step of the segmentation pipeline as shown in Figure 4 below.

[0066] Figure 4 illustrates the second segmentation step of the morphological based segmentation pipeline performed by a segmentation unit, such as the segmentation unit 104 of the system 100, according to an embodiment of the present disclosure.

[0067] Figure 4 shows an input image 402, an adjusted image 404, a denoised image 406, a blurred image 408, a binarized image 410, a closed image 412, an eroded image 414, a dilated image 416, a filled image 418, a single object image 420, a tissue region image 422, and a segmentation boundary image 424.

[0068] The morphological segmentation pipeline illustrated in Figure 4 comprises a plurality of sequential morphological operations and other image processing operations. The morphological operations accurately identify the boundary of the tissue within the input image 402 which corresponds to the tissue region. The use of the morphological operations allow geometric descriptors to be extracted accurately from the tissue region thereby improving the performance of down-stream tasks which utilize these descriptors such as quality control and estimation of tissue phenotype.

[0069] The input image 402 corresponds to the output image of the first segmentation step of the morphological based segmentation pipeline (i.e., the output image 328 shown in Figure 3).

[0070] The adjusted image 404 is generated from the input image 402 to reduce the intensity of the well border (as seen by the dark border surrounding the input image 402). The generation of the adjusted image 404 helps to improve the performance of the tissue segmentation process because the pixel intensity values of the well border lie within a similar range to the pixel intensity values of the tissue region within the input image 402. The adjusted image 404 is generated by identifying a 50 pixel border region within the input image 402 and then, for any pixel within the border region having a pixel intensity value below 0.1 (i.e., below 10% of the pixel intensity range of the input image 402), replacing the pixel with a constant value. In one implementation, the constant value is the mean pixel intensity value of the input image 402 with the standard deviation of pixel intensity values of the input image 402 added.

[0071] A denoising algorithm is applied to the adjusted image 404 to produce the denoised image 406. Any suitable denoising algorithm can be used such as a Gaussian blur, median filter, and the like. In one implementation, a non-local means approach is used with a filter strength of 100, a template window size of 50, and a search window size of 21.

[0072] The denoised image 406 is then blurred using a non-symmetrical filter to produce the blurred image 408. The use of the non-symmetrical filter acts to filter out the tissue scaffolds from the denoised image 406 by reducing the intensity of the pixels within the denoised image 406 that correspond to the tissue scaffold. As the tissue scaffold appears in the denoised image 406 as substantially horizontal lines, a non-symmetrical filter is used to filter out the tissue scaffold without removing important tissue boundary details. More particularly, the non-symmetrical filter suppresses any horizontal edges and subsequently enhances any vertical edges. The non-symmetrical filter initially convolves the denoised image 406 with a kernel of the form : is normalized before convolution such that the denoised image 406 is convolved with the normalized kernel = a)1 / ^i- The result of the convolution is then convolved with a kernel of the form :

[0073] Where eo2is normalized before convolution such that the result of convolving the denoised image 406 with is convolved with the normalized kernel eo2' to generate the blurred image 408,

[0074] The blurred image 408 is then binarized to produce the binarized image 410. A morphological closing operation is applied to the binarized image 410 to produce the closed image 412. Morphological closing comprises a dilation operation followed by an erosion operation, where both operations utilize the same structuring element. The dilation operation is the opposite of the erosion operation : the value of a pixel in the output image is set to the maximum value of all pixels within the neighborhood of the pixel within the input image (where the neighborhood is defined by the structuring element). The purpose of morphological closing is to fill (or close) small holes within the binarized image 410 whilst maintaining the geometry (shape and size) of the larger objects within the binarized image 410 such as the connected component which roughly corresponds to the tissue region. In one implementation, the morphological closing operation is performed with a square structured element of connectivity one (i.e., only the nearest neighbors are connected, diagonally connected elements are not considered neighbors).

[0075] The closing operation is followed by separate erosion and dilation operations. That is, the closed image 412 is eroded to produce the eroded image 414, and the eroded image 414 is subsequently dilated to produce the dilated image 416. Unlike morphological opening, where the structuring element is the same for both the erosion and the dilation, the structuring element used to produce the eroded image 414 is different to the structuring element used to produce the dilated image 416. The erosion and dilation operations remove small objects and lines from the closed image 412 whilst maintaining the geometry (shape and size) of the larger objects within the closed image 412 such as the connected component which roughly corresponds to the tissue region. In one implementation, the erosion and dilation operations are performed using a cross-shaped structured element with a square connectivity of one.

[0076] A filling operation is then applied to the dilated image 416 to produce the filled image 418. The largest connected component is then extracted from the filled image 418 to produce the single object image 420. The single object image 420 thus corresponds to the segmentation map of the tissue region within the input image 402. This is illustrated in the tissue region image 422 and the segmentation boundary image 424 which respectively show the tissue region extracted from the input image 402 and the boundary of the tissue region (i.e., the boundary of the segmentation of the image region within the single object image420) overlaid the input image 402. The boundary of the tissue region is shown as a white line circumscribing the tissue within the segmentation boundary image 424. As can be seen, the two step segmentation process accurately extracts the tissue region despite the presence of extraneous structures and noise within the image.

[0077] Referring once again to Figure 1, after the regions have been segmented by the segmentation unit 104, the extraction unit 106 extracts one or more features from the regions (or segmentation maps) produced by the segmentation unit 104.

[0078] Figure 5 illustrates features which are extracted from tissue regions of a segmented image according to an aspect of the present disclosure.

[0079] Figure 5 shows a tissue region 502 circumscribed by a boundary 504. Figure 5 further shows a centroid 506 of the tissue region 502, a first boundary point 508 along the boundary 504 of the tissue region 502, and a second boundary point 510 along the boundary 504 of the tissue region 502. A plurality of mid-points 512-1, 512-2, 512-3, 512- 4 are spaced between the first boundary point 508 and the second boundary point 510. Figure 5 further shows a first tissue scaffold region 514 and a second tissue scaffold region 516.

[0080] A first expanded view 518 of the first tissue scaffold region 514 shows a portion of the tissue region 502 around a first tissue scaffold 520 of the device within which the image of the artificial tissue was taken. The skilled person will appreciate that the first tissue scaffold 520 is overlaid the tissue region 502 for illustration purposes and does not form a part of the tissue region 502 extracted by a segmentation unit such as the segmentation unit 104 in Figure 1. The first expanded view 518 shows a first intersection point 522 associated with a first edge of the first tissue scaffold 520 and a second intersection point 524 associated with a second edge of the first tissue scaffold 520.

[0081] A second expanded view 526 of the second tissue scaffold region 516 shows a portion of the tissue region 502 around a second tissue scaffold 528 of the device within which the image of the artificial tissue was taken. The skilled person will appreciate that the second tissue scaffold 528 is overlaid the tissue region 502 for illustration purposes and does not form a part of the tissue region 502 extracted by a segmentation unit such as the segmentation unit 104 in Figure 1. The second expanded view 526 shows a third intersection point 530 associated with a first edge of the second tissue scaffold 528 and a fourth intersection point 532 associated with a second edge of the second tissue scaffold 528.

[0082] The tissue region 502 is extracted from an image of an artificial tissue using a segmentation approach such as that described above in relation to Figures 3 and 4 (e.g., by the segmentation unit 104 in Figure 1). As such, the tissue region 502 corresponds to a single connected component, or object, within a segmentation or labelling image. For example, the tissue region 502 corresponds to a connected region of foreground pixels or pixels assigned a value indicative of being part of the tissue region. As will be described in more detail below, the points illustrated in Figure 5 within, or proximate to, the tissue region 502 (e.g., the centroid 506, the first boundary point 508, etc.) form landmark points which are used to estimate one or more geometric features, or parameters, of the artificial tissue such as tissue length, tissue width, tissue area, etc. These features are used to form a tissue descriptor which is used in downstream tasks such as quality control and tissue phenotype encoding.

[0083] The points (e.g., the centroid 506, the first boundary point 508, etc.) are determined from the tissue region 502 by initially computing the central moments of the tissue region 502. The central moments are then used to determine the centroid 506 of the tissue region 502. The remaining points are determined by identifying the major axis of the tissue region 502. As can be seen in the example in Figure 5, the tissue region 502 is not necessarily oriented to the axes of the image— i.e., the orientation of the tissue region 502 does not necessarily align with that of the image. The orientation of the tissue region 502 is determined by computing the inertia tensor for the tissue region 502 based on the central moments. The eigenvector of the inertia tensor associated with the largest eigenvalue of the inertia tensor corresponds to the major, or principle, axis of the tissue region 502. The minor axis of the tissue region 502 is thus determined as perpendicular to the major axis of the tissue region 502. A line extending from the centroid 506 along the major axis to the boundary 504 of the tissue region 502 is then fit to identify the first boundary point 508 and the second boundary point 510.

[0084] The plurality of mid-points 512-1, 512-2, 512-3, 512-4 are then determined. The plurality of mid-points include: a first mid-point 512-1 and a fourth mid-point 512-4 associated with the first tissue scaffold region 514 and the second tissue scaffold region 516; and a second mid-point 512-2, a third mid-point 512-3 and the centroid 506. The second midpoint 512-2 and the third mid-point 512-3 are determined by identifying three evenly spaced points between the first boundary point 508 and the second boundary point 510. The skilled person will appreciate that these evenly spaced points would incorporate the centroid 506 which is separately identified in Figure 5 for illustrative purposes. As such, references to the plurality of mid-points is considered to refer to the plurality of points interposed between the first boundary point 508 and the second boundary point 510 (i.e., the plurality of mid-points 512-1, 512-2, 512-3, 512-4 shown in Figure 5 along with the centroid 506). In Figure 5, five mid-points are determined. Alternatively, between 5 and 50 mid-points are determined between the first boundary point 508 and the second boundary point 510.

[0085] The first tissue scaffold region 514 and the second tissue scaffold region 516 correspond to predetermined bounding boxes. That is, because the segmented image from which the regions shown in Figure 5 are extracted is a constant size and resolution (i.e., the segmentation process described in relation to Figures 3 and 4 results in an image of constant, or consistent, dimensions / resolution), the locations of the first tissue scaffold region 514 and the second tissue scaffold region 516 are substantially consistent across different images. Therefore, the first tissue scaffold region 514 and the second tissue scaffold region 516 are extracted from an image (e.g., the image 200 shown in Figure 2 or the input image 302 shown in Figure 3) by extracting portions of the image that lie within predetermined bounding boxes. The predetermined bounding boxes can be identified from one or more training, or example, images using either manual or automatic identification processes (e.g., image registration and feature detection). In one implementation, where the image has (height x width) dimensions of 1500px x 500px such that the tissue region extends vertically within the image, the bounding boxes have (height x width) dimensions of 400px x 500px, the bounding box of the first tissue scaffold region 514 has a top-left anchor point at image location (0,185) and the bounding box of the second tissue scaffold region 516 has a top-left anchor point at image location(0,965).

[0086] The first intersection point 522 shown within the first tissue scaffold region 514 is determined from a first reference line which is fit to the inner edge of the first tissue scaffold 520 (i.e., the lower edge of the first tissue scaffold 520 shown in Figure 5). More particularly, the first reference line is fit to the inner edge of the first tissue scaffold 520 at or near the point of inflection of the inner edge such that the first reference line approximates the tangent to the inner edge of the first tissue scaffold 520 at or near the point of inflection. The first intersection point 522 is then determined as the intercept of the first reference line with an edge of the image or image region (e.g., the y-intercept if the tissue extends vertically within the image). Once the first reference line has been fit, the gradient of the first reference line is determined and the second intersection point 524 is determined from a second reference line which has the same gradient as the first reference line (i.e., both reference lines are parallel). The second reference line is fit such that it passes through, or touches, the point of the inner edge of the first tissue scaffold 520 which is furthest from the first reference line. The intercept value (e.g., y- intercept) of the second reference line corresponds to the second intersection point 524. In practice, the second intersection point 524 approximately corresponds to the fixing point of the first tissue scaffold 520 to the well such that the absolute distance between the first intersection point 522 and the second intersection point 524 approximates the curvature, or tension, of the first tissue scaffold 520. This metric is used to determine the passive tension of the tissue scaffold 520— i.e., the estimated tension that the tissue exerts on the tissue first scaffold 520 when at rest.

[0087] Using the reference points (e.g., the plurality of mid-points 512-1, 512-2, 512-3, 512-4, the centroid 506, the first intersection point 522, etc.), one or more tissue parameters are extracted which can be subsequently used to determine a descriptor of the tissue. The tissue parameters include geometric parameters (or geometric features) such as tissue length, tissue width(s), tissue area, and tissue volume. The tissue parameters further include non-geometric parameters (or non-geometric features) such as passive tension and one or more irregularity gradings. Tissue length is a geometric feature, or parameter, of the tissue corresponding to the longitudinal length of the artificial tissue (i.e., the length of the artificial tissue's major axis). The length is determined by calculating the distance between the first boundary point 508 and the second boundary point 510. In one embodiment, the pixel-based length is converted to / j.m based on the resolution of the tissue image from which the tissue region 502 was obtained.

[0088] Tissue widths are geometric features, or parameters, of the tissue corresponding to measurements taken along the minor axis of the artificial tissue (i.e., cross-sectional measurements). A width of the artificial tissue is determined at each of the plurality of mid-points 512-1, 512-2, 512-3, 512-4 as well as at the centroid 506 thereby generating a plurality of width measurements. At a first point along the longitudinal central axis of the artificial tissue (e.g., the centroid 506), the tissue width is determining by extending a line along the minor axis of the artificial tissue (which is perpendicular to the major axis) at the first point and identifying the pair of points at which the line intersects the boundary 504. The distance between the pair of points thus corresponds to the width at the first point. Repeating this for each of the plurality of mid-points 512-1, 512-2, 512-3, 512-4 and the centroid 506 produces a plurality of widths of the artificial tissue. In one embodiment, the pixel-based length is converted to / j.m based on the resolution of the tissue image from which the tissue region 502 was obtained. In one embodiment, 5 tissue widths are determined from 5 mid-points. Alternatively, between 5 and 50 tissue widths are determined from between 5 and 50 mid-points spaced between the first boundary point 508 and the second boundary point 510.

[0089] Tissue area is a geometric feature, or parameter, corresponding to the estimated surface area of the artificial tissue as captured within the image. The tissue area is determined from the number of pixels within the tissue region 502. In one embodiment, the pixelbased area is converted to / j.m2or mm2based on the resolution of the tissue image from which the tissue region 502 was obtained.

[0090] Tissue volume is a geometric feature, or parameter, corresponding to the estimated volume of the artificial tissue. A slice volume model is used to estimate volumetric measurements from a 2-dimensional segmentation of a tissue (i.e., the tissue region 502). At a general level, the slice volume model estimates a volume of a slice of the artificial tissue at a respective location, or point, along the longitudinal axis based on an estimated cross-sectional area of the artificial tissue. The estimated cross-sectional area is determined from a width of the artificial tissue at the respective location, or point. As will be described in more detail below, the estimated cross-sectional area is determined according to a model parameterized by the width of the artificial tissue and a predetermined tissue elongation parameter.

[0091] The tissue volume is determined as a piecewise combination of local slice volumes determined at n locations, or points, distributed evenly along the longitudinal line extending between the first boundary point 508 and the second boundary point 510. The skilled person will appreciate that these locations have been omitted from Figure 5 for brevity. In one embodiment, n > 10, more particularly 10 < n < 100, and more particularly again n = 40. For the i-th slice (i.e., at the i-th point along the longitudinal line), the slice width, wifis determined by extending a line along the minor axis of the artificial tissue at the point i and calculating the width as the distance between the pair of points where the line intersects the boundary 504. The tissue volume, V , is then computed as:

[0092] The expression within brackets in Equation (1) above corresponds to a local model of the cross-sectional area of the artificial tissue at the i-th slice where a is a predetermined tissue elongation parameter. In general, the tissue elongation parameter a relates to the eccentricity of the cross-sectional shape of the artificial tissue. In one embodiment, a > 1, more particularly 1 < a < 5, and more particularly again a = 2.5. The term t, is the thickness of the i-th slice. In one embodiment, t, = where I is the length of the tissue region 502 (i.e., the distance between the first boundary point 508 and the second boundary point 510). In one embodiment, the pixel-based volume is converted to / j.m3or mm3based on the resolution of the tissue image from which the tissue region 502 was obtained.

[0093] As will be described in more detail below in relation to Figure 13, estimating the volume of an artificial tissue from a 2-dimensional image of the artificial tissue provides an efficient and discriminative encoding of the phenotype of the artificial tissue. In addition, the model described in relation to Equation (1) above allows the volume of an artificial tissue to be estimated efficiently without the need for expensive and computationally intensive volumetric measurements to be obtained (e.g., computed tomography (CT) scans, magnetic resonance imaging (MRI) scans, and the like). Indeed, any existing 2-dimensional tissue imaging system can be quickly and efficiently extended to provide volumetric estimates using the above described model.

[0094] The above geometric features are used to generate one or more tissue descriptors which encode a phenotype of the artificial tissue. The one or more tissue descriptors include an area descriptor, a volume descriptor, and a smoothness descriptor. The area and the volume descriptors correspond to the area of the tissue region 502 and the volume of the artificial tissue respectively as described in detail above. The smoothness descriptor encodes the smoothness of the boundary 504 based on the plurality of widths. That is, for a plurality of widths, W = {w ”=1, determined using the approach described above, the smoothness descriptor, S, comprises the variance of the plurality of widths determined as:

[0095] „ _ Z(ivt- lT) (4)

[0096] (n - 1) ’ where W is the mean tissue width. Alternatively, the smoothness descriptor is determined using any other suitable measurement of the variability or dispersion of the plurality of widths.

[0097] As will be described in more detail below, the one or more tissue descriptors provide a comparable geometric encoding of tissue phenotype which can be used to determine phenotypic change resulting from a drug treatment, disease model, different cell line, or other perturbation. The one or more tissue descriptors, and / or the tissue parameters from which the one or more tissue descriptors are determined, also provide a mechanism for tracking the maturation of an artificial tissue for quality control (QC) purposes.

[0098] Non-geometric features which can be extracted from the tissue regions identified by a segmentation unit include passive tension measures and tissue irregularity gradings.

[0099] The passive tension measures, or passive tension parameters, correspond to the estimated tension(s) that the artificial tissue exerts on the tissue scaffold(s) when at rest (i.e., when the artificial tissue is not being stimulated by an external stimulation source). A passive tension measure is estimated for each of the tissue scaffolds upon which the artificial tissue is attached. Alternatively, a single passive tension measure is determined for one of the tissue scaffolds. As illustrated in Figure 5, passive tension is measured from the first tissue scaffold region 514 and / or the second tissue scaffold region 516 identified by the segmentation unit 104 as described above.

[0100] For the first tissue scaffold region 514, the passive tension is computed by determining the absolute difference between the first intersection point 522 and the second intersection point 524. This absolute difference corresponds to the (vertical) distance between the anchor points of the first tissue scaffold 520 and the most extreme curvature point of the first tissue scaffold 520. The passive tension is reported in pixels and can be converted to micrometers based on the resolution of the image from which the tissue region 502 is obtained. Similarly, for the second tissue scaffold region 516, the passive tension is computed by determining the absolute difference between the third intersection point 530 and the fourth intersection point 532. As mentioned above, tissue irregularity grading (or simply irregularity grading / scoring) is another non-geometric feature which can be extracted from the tissue regions. Tissue irregularity grading is primarily used for quality control (QC) purposes and quantifies irregularities, or problems, related to the artificial tissue. Irregularities include: holes or discontinuities within the artificial tissue; irregularities in tissue shape; debris within the tissue; a broken tissue; a broken tissue scaffold; and irregularities associated with attachments of the artificial tissue to the device within which the artificial tissue is contained (e.g., attachment to the side of the well, or a lack of attachment to the tissue scaffolds). Beneficially, tissue irregularity grading can be used to supplement experimental procedures and predictions (e.g., as performed by the assay unit 110 shown in Figure 1) by providing an explanation as to variability in experimental outcomes. For example, if the phenotypic or functional response of an artificial tissue does not correspond to expected values, then the tissue irregularity grading(s) for the artificial tissue may indicate problems with the artificial tissue (e.g., abnormal attachment, hole within the artificial tissue, etc.) which may at least in part explain the unexpected or abnormal response.

[0101] Holes, or discontinuities, within the artificial tissue are identified using one or more morphological operations. In one embodiment, the number of holes within the artificial tissue is determined by calculating the Euler number, or Euler characteristic, for the tissue region 502. The Euler number corresponds to the number of objects (i.e., connected components) minus the number of holes. For a tissue region without any holes, the Euler number would be 1 because there is a single object (the tissue region) and no holes. For a tissue region with a single hole, the Euler number would be 0 because there is a single object (the tissue region) and a single hole. Similarly, for a tissue region with two holes, the Euler number would be -1 (minus one) because there is a single object (the tissue region) and two holes. The irregularity score for holes or discontinuities within the artificial tissue can thus be determined according to the Euler number of the tissue region 502. Holes within the artificial tissue are reported either as a binary value (e.g., "True" if the Euler number is less than 1 and "False" if the Euler number is equal to 1) or as a numerical value (e.g., 1 minus the Euler number provides a numerical value of the number of holes within the tissue region).

[0102] Irregularities in tissue shape, debris within the tissue, breaks in the tissue, and breaks in the tissue scaffold are identified using one or more machine learning models. More particularly, a machine learning model is trained to identify each of these irregularities from an image of the tissue (e.g., from an image of the tissue within the well region such as the well region 202 shown in Figure 2). Tissue shape irregularities are classified as either regular (normal) or irregular. Debris within the tissue is classified as either being present (i.e., there is debris within the tissue) or absent (i.e., there is no debris within the tissue). Breaks in the tissue or scaffold are identified as either present (i.e., there is at least one break in the tissue or scaffold) or absent (i.e., there are no breaks in the tissue or scaffold). As such, each of the irregularities can be identified by a binary classifier trained to predict the presence or absence of the irregularity. In one embodiment, the binary classifier trained for each irregularity is a VGG-16 convolutional neural network (CNN). Each binary classifier, or irregularity classifier, is trained on a training data set of approximately 2,000 images having approximately half of the images within each of the binary classes. For each classifier, mini-batch gradient descent with momentum is used fortraining with a batch size of 128 and momentum 0.9. Each of the above mentioned irregularities are reported as a binary value— "True" if the irregularity is present, and "False" if the irregularity is not present.

[0103] In addition to the above irregularities, irregularity grading can include one or more tissue attachment gradings indicative of a degree of attachment of the artificial tissue to the device, and / or elements thereof such as the tissue scaffold(s), within which the artificial tissue is contained. Tissue attachment gradings provide an important quantification of potentially irregular formation of an artificial tissue because the tissue attachment gradings identify irregular attachment points of the artificial tissue in relation to the device within which it is contained. In an exemplary setting, after maturation the artificial tissue is attached to the tissue scaffolds of the device but not to any other element of the device (e.g., sidewalls of the well). Attachments of the artificial tissue to other elements of the device may inhibit the functional response of the artificial tissue thus causing irregular or inaccurate phenotype readings or functional response values. As such, tissue attachment gradings help identify and quantify issues regarding the state of the artificial tissue (e.g., during maturation) which helps improve the quality of artificial tissue produced by identifying sources of systemic manufacturing variability. During an example QC process, artificial tissues identified as having tissue attachment gradings which meet certain criteria (e.g., a severe attachment to the side of the well away from the tissue scaffold) can be removed thereby allowing new artificial tissue to be grown in place.

[0104] In general, a tissue attachment grading for an image of an artificial tissue comprises three elements: (i) an indication of whether the artificial tissue is attached to an attachment site of the device; (ii) a measure or indication of the degree (or attachment severity score) of attachment of the artificial tissue to the attachment site; and (iii) a measure of the translucence of the artificial tissue at or near the attachment site. Potential attachment sites for which tissue attachment gradings are computed include a tissue scaffold, a side (or side wall) of the well near to the tissue scaffold, and a side (or side wall) of the well away from the tissue scaffold. The side of the well near to the tissue scaffold corresponds to a portion of the side of the well near to, or proximal to, the tissue scaffold (e.g., a portion of the side wall of the well over which the tissue scaffold passes, or over which the tissue scaffold overlaps). The side of the well away from the tissue scaffold corresponds to a portion of the side of the well which is away from, or distal to, the tissue scaffold. A tissue attachment grading is computed for one or more of the potential attachment sites at each "corner" of the artificial tissue such that each artificial tissue has four attachment grading labels for each attachment site (e.g., a top-left, top-right, bottom-left, bottomright score for attachment to the tissue scaffold, a top-left, top-right, bottom-left, bottomright score for attachment to the side of the well near to the tissue scaffold, etc.). This is illustrated in Figure 6.

[0105] Figure 6 shows the decomposition of a well-cropped image of an artificial tissue for tissue attachment grading according to an embodiment of the present disclosure.

[0106] Figure 6 shows an image 602 of an artificial tissue which corresponds to an image of an artificial tissue cropped to the well region (as identified by a segmentation unit such as the segmentation unit 104 of Figure 1 as described above). The image 602 is split into four quadrant images: an upper-left quadrant 604, an upper-right quadrant 606, a lower- left quadrant 608, and a lower-right quadrant 610. Figure 6 further shows an indication of attachment sites including a first attachment site 612 at the tissue scaffold, a second attachment site 614 at the side of the well near to the tissue scaffold, and a third attachment site 616 at the side of the well away from the tissue scaffold.

[0107] The four quadrant images all have the same width and height corresponding to half of the width and height of the image 602. Each quadrant is transformed so that the artificial tissue within each quadrant has the same orientation. That is, the upper-right quadrant is extracted from the image 602 and then horizontally mirrored to generate the upper-right quadrant 606 shown in Figure 6; the lower-left quadrant is extracted from the image 602 and vertically mirrored to generate the lower-left quadrant 608 shown in Figure 6; and the lower-right quadrant is extracted from the image 602 and mirrored both horizontally and vertically to generate the lower-right quadrant 610 shown in Figure 6.

[0108] Rather than determine a single attachment grading for the artificial tissue within the image 602 (for a respective attachment site), four attachment gradings are determined from the four quadrant images for the respective attachment site (e.g., the first attachment site 612, the second attachment site 614, etc.). Beneficially, this helps provide a greater level of detail regarding the tissue's attachment to the device whilst also enabling models for predicting tissue attachment grading to be more efficiently trained, as discussed in more detail below.

[0109] The tissue attachment grading obtained from an image of a tissue attachment region (e.g., a quadrant image such as any of those shown in Figure 6) can be represented hierarchically, as shown in Figure 7. In the following description in relation to Figures 7 and 9, the skilled person will appreciate that references in general to tissue attachment gradings encompasses any of the tissue attachment gradings (i.e., any of the tissue attachment sites) described above

[0110] Figure 7 shows a hierarchical tissue attachment grading 700 according to embodiments of the present disclosure.

[0111] The hierarchical tissue attachment grading 700 is associated with an input image 702 of a region of an artificial tissue (e.g., a quadrant image such as any of those shown in Figure 6). The hierarchical tissue attachment grading 700 comprises three levels: an attachment output level 704, a severity output level 706, and a translucence output level 708. The attachment output level 704 comprises either an attachment not present label 710 or an attachment present label 712. The severity output level 706 comprises either a minor attachment label 714, a moderate attachment label 716, or a severe attachment label 718. In some embodiments, the severity output level 706 comprises an attachment not present label 720. The translucence output level 708 comprises either a not translucent label 722 or a translucent label 724. In some embodiments, the translucence output level 708 comprises an attachment not present label 726.

[0112] The label assigned at the attachment output level 704 can be represented by labels {no attachment, attachment} (e.g., the attachment not present label 710 or the attachment present label 712) or via a 2-valued indicator vector where the first value is used to identify no attachment (e.g., the vector [1,0]) and the second value is used to identify attachment (e.g., the vector [0,1]). The labels assigned at the severity output level 706 can be represented by labels {none, minor, moderate, severe} (e.g., the attachment not present label 720, the minor attachment label 714, the moderate attachment label 716, or the severe attachment label 718) or via a 4-valued indicator vector where the first value is used to identify no attachment (e.g., the vector [1,0, 0,0]), the second value is used to identify minor attachment (e.g., the vector [0,1, 0, 0]), the third value is used to identify moderate attachment (e.g., the vector [0,0, 1,0]), and the fourth value is used to identify severe attachment (e.g., the vector [0,0, 0,1]). The label assigned at the translucence output level 708 can be represented by labels {no attachment, not translucent, translucent} (e.g., the attachment not present label 726, the not translucent label 722, or the translucent label 724) or via a 3-valued indicator vector where the first value is used to identify no attachment (e.g., the vector [0,1, 0]), the second value is used to identify not translucent (e.g., the vector [0,1, 0]), and the third value is used to identify translucent (e.g., the vector [0,0,1]).

[0113] The hierarchical tissue attachment grading 700 can be represented as a combination of labels (e.g., {attachment, minor, not translucent} for a non-translucent tissue having a minor attachment at the respective tissue attachment site), a numerical value representing each of the seven different possible outcomes, or an indicator vector representing each of the different possible outcomes (e.g., the vector [1, 0, 0, 0, 0, 0, 0 ] represents the no attachment outcome shown by the " = 0" path in Figure 7, the vector [0, 0, 0, 1, 0, 0, 0 ] represents the moderate attachment not translucent outcome shown by the " = 3" path in Figure 7, and the vector [0, 0, 0, 0, 0, 0, 1 ] represents the severe attachment translucent outcome shown by the " = 6" path in Figure 7.

[0114] The attachment grading is determined using a grading model. In general, the grading model is configured or trained to receive an input image (e.g., the input image 702), or a set of features extracted therefrom, and estimate an attachment grading based on the input image. In one embodiment, the grading model comprises a hierarchy of classifiers such that a classifier is associated with each output level in the hierarchical tissue attachment grading 700. That is, an attachment classifier determines a label for the attachment output level 704 based on the input image 702, a severity classifier determines a label for the severity output level 706 based on the input image 702, and a translucence classifier determines a label for the translucence output level 708 based on the input image 702. The attachment grading is then determined by combining the outputs of each of the classifiers within the hierarchy of classifiers.

[0115] As stated above, one or more feature extraction processes can be applied to the input image 702 such that features obtained from the one or more feature extraction processes are used as input to the grading model (e.g., each of the classifiers within the hierarchy of classifiers) during both the training stage and the inference stage. Example feature extraction processes include principal components analysis (PCA), bag-of-words models, histogram of oriented gradients (HOG) features, speeded up robust features (SURF), and local binary pattern (LBP) features.

[0116] The attachment classifier assigns either the attachment not present label 710 or the attachment present label 712 to the input image 702 (or features extracted therefrom). As such, the attachment classifier is a binary classifier trained to predict the presence of absence of an attachment within an input image. In one embodiment, the attachment classifier determines a probability, or score, associated with the presence of an attachment within an input image. The probability, or score, is then converted to a label based on a thresholding (e.g., if the probability is greater than 0.5, then the attachment present label 712 is assigned, otherwise the attachment not present label 710 is assigned). The attachment classifier is any suitable binary machine learning classifier such as a naive Bayes classifier, a support vector machine (SVM), a multilayer perceptron, a decision tree, a k-nearest neighbor (kNN) classifier, or the like.

[0117] The severity classifier assigns either the minor attachment label 714, the moderate attachment label 716, or the severe attachment label 718 to the input image 702 (or features extracted therefrom). In one embodiment, the severity classifier may also assign the attachment not present label 720 to the input image 702 (i.e., in place of either the minor attachment label 714, the moderate attachment label 716, or the severe attachment label 718). As such, the severity classifier is a multi-class classifier trained to predict the severity or degree of attachment of an artificial tissue to a tissue scaffold (as represented within an input image). In one embodiment, the severity classifier determines a probability, or score, for each of the possible outputs— minor, moderate, severe (and optionally no attachment)— and assigns a label based on the most probable output (i.e., the output with highest probability or score). The severity classifier is any suitable multiclass classifier such as a naive Bayes classifier, a multilayer perceptron, a decision tree, a kNN classifier, or the like.

[0118] The translucence classifier assigns either the not translucent label 722 or the translucent label 724 to the input image 702 (or features extracted therefrom). In one embodiment, the translucence classifier may also assign the attachment not present label 726 to the input image 702 (i.e., in place of the not translucent label 722 or the translucent label 724). As such, the translucence classifier is a binary (or multi-class) classifier trained to predict the translucence of an artificial tissue at or near a tissue scaffold (as represented within an input image). In one embodiment, the translucence classifier determines a probability, or score, for each of the possible outputs— translucent, not translucent (and optionally no attachment)— and assigns a label based on the most probable output. The translucence classifier is any suitable binary (or multi-class) classifier such as a naive Bayes classifier, a support vector machine (SVM), a multilayer perceptron, a decision tree, a k-nearest neighbor (kNN) classifier, or the like.

[0119] Each classifier within the hierarchy of classifiers is trained on a training data set of labelled images with labels related to a specific attachment site. For example, a first attachment classifier is trained on a training dataset of images with labels relating to the attachment of the artificial tissue shown within each image to the tissue scaffold shown within each image (e.g., each label comprises an attachment indicator relating to whether the artificial tissue is attached to the tissue scaffold). A first severity classifier is trained on the training dataset of images with labels relating to the degree of attachment of the artificial tissue shown within each image to the tissue scaffold (e.g., each label comprises an attachment severity score relating to the degree of attachment of the artificial tissue to the tissue scaffold). A first translucence classifier is trained on the training dataset of images with labels relating to the translucence of the tissue shown within each image (e.g., each label comprises a translucence score relating to the translucence of the artificial tissue at or near to the tissue scaffold). The training dataset thus comprises a plurality of images and a plurality of labels each of which being associated with a corresponding image of the plurality of images.

[0120] In one implementation, a hierarchy of classifiers is trained for each attachment site. Each classifier within the hierarchy of classifiers comprises a support vector machine (SVM) with a Kullback-Leibler (K-L) divergence based kernel. During training and inference, a feature extraction pipeline is used to convert a quadrant image to a feature vector. Uniform local binary patterns are extracted from the quadrant image with a cell size of 16x16, a radius of 1, and 8 neighbor set points. The set of local binary patterns for each image are then converted to a feature vector by generating a histogram from the set of local binary patterns (with the number of bins set to the number of unique local binary patterns + 1). Each image is thus represented by a feature vector corresponding to the histogram of occurrences of each local binary pattern within the image. Because these feature vectors may be considered a distribution of local binary patterns within the image, the K-L divergence based kernel is used for the SVM. The hyperparameters of the SVM are determined using a standard grid search based method. Each of the classifiers within the hierarchy of classifiers are trained on a training dataset comprising 2,000 images obtained from images of 500 artificial tissues such that each image within the training dataset corresponds to a quadrant image. Each image within the training dataset is associated with a label vector within a corresponding plurality of labels. The label vector comprises an attachment score, a severity score, and a translucence score for each of the three attachment sites (at the tissue scaffold, at the portion of the side of the well near to the tissue scaffold, and at the portion of the side of the well away from the tissue scaffold) such that each label within the plurality of labels comprises a 9-dimensional vector. The classifiers are then trained to predict a given label within the label vector. For example, an attachment classifier for the tissue scaffold attachment site is trained to predict the label within the label vector associated with the attachment score for the tissue scaffold attachment site. The outputs of each classifier of the hierarchy of classifiers are combined to produce the attachment grading. As stated above, the outputs can be combined to generate a label vector (e.g., {attachment, minor, not translucent}), a numerical representation (e.g., "1"), or an indicator vector (e.g., [0,1,0, 0, 0, 0, 0]).

[0121] Figure 8 shows cropped regions of three artificial tissues having different predicted tissue attachment gradings according to embodiments of the present disclosure.

[0122] Figure 8 shows a first attachment image 802, a second attachment image 804, and a third attachment image 806. Each of the attachment images is obtained from a corresponding image of an artificial tissue cropped to a region of the image comprising a side of the well near to a tissue scaffold. The skilled person will appreciate that the images have been cropped in this way to aid understanding and provide a clearer view of the tissue attachment site. In the examples shown in Figure 8, the attachment site corresponds to the side of the well near the tissue scaffold (e.g., the second attachment site 614 shown in Figure 6).

[0123] The first attachment image 802 shows a predicted minor attachment of the artificial tissue to the side of the well with the artificial tissue being graded as translucent at or near the attachment site. As such, the tissue attachment grading forthe first attachment image 802 can be expressed as {attached, minor, translucent}, by the numerical value "2", or by the indicator vector [0, 0, 1, 0, 0, 0, 0] .

[0124] The second attachment image 804 shows a predicted moderate attachment of the artificial tissue to the side of the well with the artificial tissue being graded as not translucent at or near the attachment site. As such, the tissue attachment grading for the second attachment image 804 can be expressed as {attached, moderate, not translucent}, by the numerical value "3", or by the indicator vector [0, 0, 0, 1, 0, 0, 0] ..

[0125] The third attachment image 806 shows a predicted severe, or strong, attachment of the artificial tissue to the side of the well with the artificial tissue being graded as translucent at or near the attachment site. As such, the tissue attachment grading for the third attachment image 806 can be expressed as {attached, severe, translucent} , by the numerical value "6", or by the indicator vector [0, 0, 0, 0, 0, 0, 1] ..

[0126] Automatically computing the tissue attachment grading for an attachment site and a sub-region of the artificial tissue (e.g., a quadrant image such as those shown in Figure 6) provides an accurate, and informative, location-specific measurement of a potential irregularity of the physiological state of an artificial tissue. This automation helps improve the reliability and robustness of the measurements reported whilst reducing the amount of manual input required thereby reducing user-induced variability. Tissue attachment gradings can thus be used to inform quality control procedures (e.g., inclusion or exclusion of an artificial tissue) and / or be used to provide supplemental analysis to help explain experimental variability.

[0127] As an alternative to the hierarchy of classifiers described above, the attachment grading for a respective attachment site is determined using a single trained multi-task learning (MTL) network as shown in Figure 9A.

[0128] Figure 9A shows a multi-task learning (MTL) network 900 for tissue attachment grading according to an aspect of the present disclosure.

[0129] The MTL network 900 comprises an input network 902, a shared dense network 904, and a plurality of output networks including an attachment network 906-1, a severity network 906-2, a translucence network 906-3, and an auxiliary network 906-4. The MTL network 900 receives an input image 908 and produces a plurality of outputs including an attachment indicator 910-1, a severity indicator 910-2, a translucence indicator 910-3, and a grading 910-4. In an embodiment, the plurality of output networks further include a minor severity network 906-5 which produces a minor severity indicator 910-5. The architecture of the input network 902 is described in more detail in relation to Figures 9B and 9C below. The architecture of the shared dense network 904 shown in Figure 9A comprises a 2-dimensional global average pooling layer 912, a first dense block 914-1, and a second dense block 914-2. The first dense block 914-1 and the second dense block 914-2 share the same dense block architecture which is illustrated in relation to the first dense block 914-1 in Figure 9A. The dense block architecture (shown in relation to the first dense block 914-1) comprises a dense layer 916, an exponential linear unit (ELU) activation layer 918, and a dropout layer 920. The skilled person will appreciate that the dense blocks shown in Figure 9A all share this same architecture but may have different hyperparameters (e.g., different dropout rate or different numbers of nodes within the dense layer). The plurality of output networks all share a common architecture, and the architecture of the output networks is shown in Figure 9A in relation to the attachment network 906-1. The output network architecture comprises a third dense block 914-3, a fourth dense block 914-4, a dense layer 922, and a softmax layer 924. As stated previously, the architectures of the third dense block 914-3 and the fourth dense block 914-4 are the same as the dense block architecture described above in relation to the first dense block 914-1. The skilled person will appreciate that, whilst the architecture of each output network is the same, the hyperparameters for each output network may vary. The MTL network 900 is trained to predict multiple outputs from a single input thereby enabling the MTL network 900 to take advantage of the provision of relational information to improve grading accuracy through updating of the shared dense network 904. The input image 908 provided to the MTL network 900 comprises a portion of an artificial tissue (e.g., a quadrant as described above in relation to Figure 6). In one example implementation, the input image 908 is a grayscale image having dimensions of 775px x 175px. As stated previously, plurality of outputs include the attachment indicator 910-1, the severity indicator 910-2, the translucence indicator 910-3, the grading 910-4, and, in some embodiments, the minor severity indicator 910-5.

[0130] The attachment indicator 910-1 (alternatively referred to as an attachment indicator vector, an attachment output vector, an attachment score, or an attachment score vector) comprises a prediction of whether there is an attachment of a portion of an artificial tissue to an attachment site within the input image 908. The attachment indicator 910-1 may be a 2-dimensional vector, where the first value of the vector comprises a probability of an attachment being present and the second value of the vector comprises a probability of no attachment being present. Because the attachment indicator 910-1 is generated by the softmax layer of the attachment network 906-1, the 2-dimensional vector is normalized such that the probabilities sum to one. The attachment indicator 910-1 may be converted to an attachment label by assigning the label which corresponds to the maximum probability value within the attachment indicator 910-1 (e.g., assigning the label {no attachment} if the first value of the vector is the maximum).

[0131] The severity indicator 910-2 (alternatively referred to as a severity indicator vector, a severity output vector, a severity score, or a severity score vector) comprises a prediction of the severity, or degree, of attachment of a portion of an artificial tissue to an attachment site (as captured within the input image 908). The severity indicator 910-2 may be a 4- dimensional vector, where the first value of the vector comprises a probability of no attachment being present, the second value of the vector comprises a probability of a minor attachment being present, the third value of the vector comprises a probability of a moderate attachment being present, and the fourth value of the vector comprises a probability of a major attachment being present. Because the severity indicator 910-2 is generated by the softmax layer of the severity network 906-2, the 4-dimensional vector is normalized such that the probabilities sum to one. The severity indicator 910-2 may be converted to a severity label by assigning the label which corresponds to the maximum probability value within the severity indicator 910-2 (e.g., assigning the label {major} if the fourth value of the vector is the maximum). The translucence indicator 910-3 (alternatively referred to as a translucence indicator vector, a translucence output vector, a translucence score, or a translucence score vector) comprises a prediction of the translucence a portion of an artificial tissue at or near an attachment site (as captured within the input image 908). The translucence indicator 910-3 may be a 3-dimensional vector, where the first value of the vector comprises a probability of no attachment being present, the second value of the vector comprises a probability of the portion of the artificial tissue being translucent, and the third value of the vector comprises a probability of the portion of the artificial tissue not being translucent. Because the translucence indicator 910-3 is generated by the softmax layer of the translucence network 906-3, the 3-dimensional vector is normalized such that the probabilities sum to one. The translucence indicator 910-3 may be converted to a translucence label by assigning the label which corresponds to the maximum probability value within the translucence indicator 910-4 (e.g., assigning the label {translucent} if the second value of the vector is the maximum).

[0132] The grading 910-4 (alternatively referred to as a grading vector, a grading output vector, or an auxiliary output vector) comprises a prediction of the overall attachment grading of a portion of an artificial tissue to an attachment site (as captured within the input image 908). The grading 910-4 may be a 7-dimensional vector, where the first value of the vector comprises a probability of no attachment being present, the second value of the vector comprises a probability of the portion of the artificial tissue being translucent with a minor attachment, the third value of the vector comprises a probability of the portion of the artificial tissue not being translucent with a minor attachment, the fourth value of the vector comprises a probability of the portion of the artificial tissue being translucent with a moderate attachment, the fifth value of the vector comprises a probability of the portion of the artificial tissue not being translucent with a moderate attachment, the sixth value of the vector comprises a probability of the portion of the artificial tissue being translucent with a major attachment, and the seventh value of the vector comprises a probability of the portion of the artificial tissue not being translucent with a major attachment. Because the grading 910-4 is generated by the softmax layer of the auxiliary network 906-4, the 7-dimensional vector is normalized such that the probabilities sum to one. The grading 910-4 may be converted to a grading label by assigning the label which corresponds to the maximum probability value within the grading 910-4 (e.g., assigning the label {attachment, moderate, not translucent} if the fifth value of the vector is the maximum). Advantageously, including the auxiliary network 906-4 (and the grading 910-4) in the MTL network 900 helps to improve the training of the MTL network 900, particularly the shared dense network 904. Once the MTL network 900 is trained, the grading 910-4 may not be used and instead the hierarchical grading for an input image determined from each of the other outputs described above.

[0133] The minor severity indicator 910-5 (alternatively referred to as a minor severity indicator vector, a minor severity output vector, a minor severity score, or a minor severity score vector) comprises a prediction of whether a portion of an artificial tissue has a minor attachment or a normal (i.e., greater than minor) attachment (as captured within the input image 908). The minor severity indicator 910-5 may be a 3-dimensional vector, where the first value of the vector comprises a probability of no attachment being present, the second value of the vector comprises a probability of the portion of the artificial tissue having a minor attachment, and the third value of the vector comprises a probability of the portion of the artificial tissue having a normal attachment. Because the minor severity indicator 910-5 is generated by the softmax layer of the minor severity network 906-5, the 3-dimensional vector is normalized such that the probabilities sum to one. The minor severity indicator 910-5 may be converted to a minor severity label by assigning the label which corresponds to the maximum probability value within the minor severity indicator 910-5 (e.g., assigning the label {no attachment} if the first value of the vector is the maximum).

[0134] The input network 902 comprises a convolutional neural network which is configured to receive the input image 908 and output a feature vector to the shared dense network 904.

[0135] Figure 9B shows the architecture of the input network 902 of the MTL network 900 shown in Figure 9A according to an embodiment of the present disclosure.

[0136] The input network 902 comprises an entry flow network 926, a plurality of middle flow networks including a first middle flow network 928-1, a second middle flow network 928-2, a third middle flow network 928-3, and a fourth middle flow network 928-4, and an exit flow network 930. An input image (e.g., the input image 908 shown in Figure 9A) is received by the entry flow network 926 and a feature vector is output from the exit flow network 930 (e.g., to the shared dense network 904 shown in Figure 9A). The entry flow network 926 comprises a first convolution block 932-1, a first 2-dimensional max pooling layer 934, a second convolution block 932-2, a third convolution block 932-3, a second 2- dimensional max pooling layer 936, a fourth convolution block 932-4, a fifth convolution block 932-5, and a third 2-dimensional max pooling layer 938. The architecture of all the convolution block (e.g., the first convolution block 932-1, the second convolution block 932-2, etc.) are the same, although the skilled person will appreciate that the hyperparameters may change between blocks. The architecture of the convolution block is shown in Figure 9B in relation to the architecture of the first convolution block 932-1 which comprises a 2-dimensional separable convolution layer 940, a batch normalization layer 942, and an exponential linear unit (ELU) activation layer 944. The architecture of the plurality of middle flow networks is described in more detail in relation to Figure 9C below. The exit flow network 930 comprises a sixth convolution block 932-6, a seventh convolution block 932-7, a 2-dimensional separable convolution layer 946, and an ELU activation layer 948.

[0137] An input image is provided to the first convolution block 932-1 of the entry flow network 926. The 2-dimensional separable convolution layer 940 of the first convolution block 932-1 has 128 filters with a filter size of (7 x 7), the batch normalization layer 942 is a normalization layer that transforms the output of the 2-dimensional separable convolution layer 940 such that the mean is approximately 0 and the standard deviation is approximately 1, and the ELU activation layer 944 applies an exponential linear activation function to the output of the batch normalization layer 942 with the scale set as a = 0.01. The output of the first convolution block 932-1 (i.e., the output of the ELU activation layer 944) is provided to the first 2-dimensional max pooling layer 934 which applies a max pooling operation (i.e., downsampling) with a pool size of (2,2). The output of the first 2-dimensional max pooling layer 934 is provided to the second convolution block 932-2 the output of which is provided to the third convolution block 932-3. As stated above, the second convolution block 932-2 and the third convolution block 932-3 share the same architecture as the first convolution block 932-1. The hyperparameters for these blocks are also the same but with the filter size for the 2-dimensional separable convolution layer of the second convolution block 932-2 being (1 x 1) and the filter size for the 2-dimensional separable convolution layer of the third convolution block 932-3 being (3 x 3). The output of the third convolution block 932-3 is provided to the second 2-dimensional max pooling layer 936 which applies max pooling with a pool size of (2,2). The output of the second 2-dimensional max pooling layer 936 is provided to the fourth convolution block 932-4 the output of which is provided to the fifth convolution block 932-5. The fourth convolution block 932-4 and the fifth convolution block 932-5 share the same architecture as the first convolution block 932-1. The hyperparameters for these blocks are also the same but with the filter size for the 2-dimensional separable convolution layer of the fourth convolution block 932-4 being (1 x 1) and the filter size for the 2-dimensional separable convolution layer of the fifth convolution block 932-5 being (3 x 3). The output of the third convolution block 932-3 is provided to the third 2- dimensional max pooling layer 938 which applies max pooling with a pool size of (2,2). The output of the entry flow network 926 (i.e., the output of the third 2-dimensional max pooling layer 938) is provided to the first middle flow network 928-1 of the plurality of middle flow networks.

[0138] Figure 9C shows the architecture of each of the plurality of middle flow networks of the input network 902 shown in Figure 9B according to an embodiment of the present disclosure.

[0139] Each middle flow network comprises a residual block 950, a max pooling block 952, a plurality of inception blocks including a first inception block 954-1, a second inception block 954-2, a third inception block 954-3, and a fourth inception block 954-4, and a dilation block 956. The input to the middle flow network is provided each of the above described blocks and the block outputs are provided to a concatenation block 958. The output of the middle flow network corresponds to the output of the concatenation block 958.

[0140] The residual block 950 comprises a 2-dimensional separable convolution layer with 128 filters of size (1 x 1). The max pooling block 952 comprises a 2-dimensional max pooling layer 960 and a 2-dimensional separable convolution layer 962. The 2-dimensional max pooling layer 960 applies max pooling with a pool size of (2 x 2) and the output of the 2-dimensional max pooling layer 960 is provided to the 2-dimensional separable convolution layer 962 which comprises 128 filters of size (1 x 1).

[0141] The plurality of inception blocks (e.g., the first inception block 954-1, the second inception block 954-2, etc.) act in parallel to extract multi-level features from the input. Each of the inception blocks share same the same architecture but with different hyperparameters for some of the layers. In Figure 9C, the common architecture is shown in relation to the first inception block 954-1. The fist inception block 954-1 comprises a first convolution block 964-1, a second convolution block 964-2, a third convolution block 964-3, and a final 2-dimensional separable convolution layer 966. The first convolution block 964-1 comprises a 2-dimensional separable convolution layer 968, a batch normalization layer 970, and an ELU activation layer 972 with the scale set as a = 0.01. The second convolution block 964-2 and the third convolution block 964-3 share the same architecture as the first convolution block 964-1 but with different hyperparameters for the 2-dimensional separable convolution layers.

[0142] For the first inception block 954-1, the 2-dimensional separable convolution layer 968 of the first convolution block 964-1 has 128 filters of size (1 x 1), the 2-dimensional separable convolution layer of the second inception block 964-2 has 128 filters of size (3 x 3), the 2-dimensional separable convolution layer of the third inception block 964-3 has 128 filters of size (1 x 1), and the final 2-dimensional separable convolution layer 966 has 128 filters of size (3 x 3). Beneficially, the use of the (1 x 1) convolution layers enables more efficient processing by reducing the dimensionality of the input data.

[0143] For the second inception block 954-2, the 2-dimensional separable convolution layer of the first convolution block has 128 filters of size (1 x 1), the 2-dimensional separable convolution layer of the second inception block has 128 filters of size (5 x 5), the 2-dimensional separable convolution layer of the third inception block has 128 filters of size (1 x 1), and the final 2-dimensional separable convolution layer has 128 filters of size (5 x 5).

[0144] For the third inception block 954-3, the 2-dimensional separable convolution layer of the first convolution block has 128 filters of size (1 x 1), the 2-dimensional separable convolution layer of the second inception block has 128 filters of size (7 x 7), the 2-dimensional separable convolution layer of the third inception block has 128 filters of size (1 x 1), and the final 2-dimensional separable convolution layer has 128 filters of size (7 X 7).

[0145] For the fourth inception block 954-4, the 2-dimensional separable convolution layer of the first convolution block has 128 filters of size (1 x 1), the 2-dimensional separable convolution layer of the second inception block has 128 filters of size (9 x 9), the 2-dimensional separable convolution layer of the third inception block has 128 filters of size (1 x 1), and the final 2-dimensional separable convolution layer has 128 filters of size (9 x 9).

[0146] The concatenation block 958 comprises a plurality of convolution layers 974-1 to 974-7 each of which comprises a 2-dimensional separable convolution layer with 128 filters of size (2 x 2) and a dilated convolution with a dilation rate of 1. Beneficially, the use of the sequential dilated convolutions within the concatenation block 958 enables the middle flow network to have a larger receptive field (i.e., by "skipping" pixels thereby enabling the convolution to cover a larger area of the input) without increasing the number of parameters thereby providing an efficient increase in receptive field without an increase in the complexity of the parameter space to be searched during training.

[0147] The concatenation block 958 concatenates the outputs of the residual block 950, the max pooling block 952, the first inception block 954-1, the second inception block 954-2, the third inception block 954-3, the fourth inception block 954-4, and the dilation block 956. The concatenation block 958 comprise a concatenation layer 976 which concatenates all of the inputs (i.e., the outputs described above), a batch normalization layer 978, and an ELU activation layer 980 with the scale set as a = 0.01. Referring once again to Figure 9B, each of the plurality of middle flow networks (e.g., the first middle flow network 928-1, the second middle flow network 928-2, etc.) have the same architecture as described in relation to Figure 9C. The output of the final middle flow network —the fourth middle flow network 928-4— is provided to the exit flow network 930.

[0148] In the exit flow network 930, the 2-dimensional separable convolution layers of the sixth convolution block 932-6 and the seventh convolution block 932-7, as well as the 2-dimensional separable convolution layer 946, have 256 filters of size (7 x 7). The ELU activation layers of the sixth convolution block 932-6 and the seventh convolution block 932-7, as well as the ELU activation layer 948, have the scale set as a = 0.01.

[0149] Referring once again to Figure 9A, the output of the input network 902 (i.e., the output of the ELU activation layer 948 of the exit flow network 930 shown in Figure 9B) is provided to the shared dense network 904. The 2-dimensional global average pooling layer 912 of the shared dense network 904 applies a global average pooling operation to the input to the shared dense network 904. The dense layer 916 of the first dense block 914-1 of the shared dense network 904 is a densely connected layer with 256 units. The ELU activation layer 918 then applies an ELU activation function with the scale set as a = 0.01 and the dropout layer 920 applies dropout to the output of the ELU activation layer 918 with a dropout rate of 0.8. The second dense block 914-2 of the shared dense network 904 has the same architecture and hyperparameters as the first dense block 914-1.

[0150] The output of the shared dense network 904 is then provided to each of the plurality of output networks. The architecture and hyperparameters of each of the plurality of output networks are the same but with differing numbers of units in the final dense layer (e.g., the dense layer 922 shown in relation to the attachment network 906-1). For each of the plurality of output networks, the dense layer of the first dense block has 256 units, the ELU activation function of the first dense block has the scale set as a = 0.01, and the dropout rate of the first dense block is set as 0.6. For each of the plurality of output networks, the dense layer of the second dense block has 512 units, the ELU activation function of the second dense block has the scale set as a = 0.01, and the dropout rate of the second dense block is set as 0.8. For the attachment network 906-1, the dense layer 922 has 2 units because the attachment network 906-1 predicts whether there is an attachment present or if there is no attachment present. For the severity network 906-2, the dense layer has 4 units because the severity network 906-3 predicts either no attachment, minor attachment, moderate attachment, or major attachment. For the translucence network 906-3, the dense layer has 3 units because the translucence network 906-3 predicts either no attachment, translucent, or not translucent. For the auxiliary network 906-4, the dense layer 922 has 7 units because the auxiliary network 906-4 predicts whether there is no attachment, a minor attachment with a translucent tissue, a minor attachment with a not translucent tissue, a moderate attachment with a translucent tissue, a moderate attachment with a not translucent tissue, a major attachment with a translucent tissue, or a major attachment with a not translucent tissue. For the minor severity network 906-5, the dense layer has 3 units because the minor severity network 906-5 predicts either no attachment, a minor attachment, or a normal attachment (i.e., an attachment which is more severe, or stronger, than a minor attachment).

[0151] In one implementation, the MTL network 900 is trained using a training dataset of 2,000 quadrant images with approximately 286 images (quadrants) for each of the 7 possible labels (e.g., labels 0 to 6 shown in Figure 7). The images were assigned a label by expert evaluation. The training dataset is split into an 80 / 20 split for training / validation. The MLT network 900 is trained using minibatch gradient descent with a batch size of 128 and an ADAM. The ADAM solver has an initial learning rate of le-3 with early stopping based on validation loss.

[0152] The outputs of a trained MTL network are used to estimate a predicted tissue attachment grading for an input image (e.g., a quadrant image such as those shown in Figure 6). The tissue attachment grading is indicative of a degree of attachment of a portion of an engineered or artificial tissue to an attachment site (as captured within the input image). In one embodiment, the auxiliary network 906-4 is used during the training phase to improve the robustness and accuracy of the predictions produced by the other output networks (e.g., the attachment network 906-1, the severity network 906-2, etc.). The predicted tissue attachment grading is then determined from the outputs of the other output networks. More particularly, the predicted attachment grading may be determined by combining the prediction of the degree of attachment (or severity / size of attachment) within the severity indicator 910-2 with the translucence prediction within the translucent indicator 910-3.

[0153] Referring once again to Figure 1, the features and / or tissue descriptors generated by the extraction unit 106 (as described above) are subsequently used by the quality control (QC) unit 108 and / or the assay unit 110. The operations of each of these units are described in turn below.

[0154] At a general level, the QC unit 108 determines the state of an artificial tissue (e.g., the physiological or morphological state) at a given time point and can track the development and / or change in state of the artificial tissue over a plurality of predetermined time points (e.g., 1 day post-seeding, 2 days post-seeding, 7 days post-seeding, etc.). At a predetermined time point, the QC unit 108 utilizes one or more tissue parameters obtained from one or more image segments of an image of the artificial tissue (taken at the predetermined time point) to generate a QC report. The one or more tissue parameters are determined by the extraction unit 106 from image segments obtained from the segmentation unit 104. The QC report provides an indication of the state of the artificial tissue at the predetermined time point as encoded within the one or more tissue parameters. The QC report also provides an indication of the change in state of the artificial tissue over the plurality of predetermined time points.

[0155] Additionally, or alternatively, the QC report comprises a tissue quality alert when at least one of the tissue parameters does not meet a predetermined quality control threshold. For example, a quality control threshold may require that the tissue attachment grading scores for each of the tissue's attachment points (i.e., the four corners of the artificial tissue) are to be above a threshold value at a certain time point (e.g., 49 days post-seeding). In one embodiment, the QC unit 108, or another unit of the system 100, causes a command to be transmitted to the bioreactor 102 based on the tissue quality alert. The bioreactor 102 may be configured to process the command (e.g., by shutting down the device containing the artificial tissue which fails to meet the quality control threshold, or issuing an alert highlighting that there is a quality issue associated with the artificial tissue which fails to meet the quality control threshold). Additionally, or alternatively, the command is transmitted to another device (e.g., a computing device, a display device, or the like) to alert a user of the system that the quality control threshold has or has not been met.

[0156] Figures 10-12 illustrate the QC report process over the maturation period of an artificial tissue.

[0157] Figure 10 shows a sequence of images of an artificial tissue during a maturation period according to an embodiment of the present disclosure.

[0158] Figure 10 shows a first image 1002 of the artificial tissue taken at a first predetermined time point during the maturation period, a second image 1004 of the artificial tissue taken at a second predetermined time point during the maturation period, a third image 1006 of the artificial tissue taken at a third predetermined time point during the maturation period, a fourth image 1008 of the artificial tissue taken at a fourth predetermined time point during the maturation period, and a fifth image 1010 of the artificial tissue taken at a fifth predetermined time point during the maturation period.

[0159] In the example shown in Figure 10 the predetermined time points correspond to 1 day post-seeding, 2 days post-seeding, 7 days post-seeding, 14 days post-seeding, and 49 days post-seeding. As such, the first image 1002 was obtained 1 day post-seeding, the second image 1004 was obtained 2 days post-seeding, the third image 1006 was obtained 7 days post-seeding, the fourth image 1008 was obtained 14 days post-seeding, and the fifth image 1010 was obtained 49 days post-seeding.

[0160] As can be seen, the morphological state of the artificial tissue evolves over the predetermined time points. This change of state is encoded within a QC report. Thus, at each of the predetermined time points, one or more tissue parameters are extracted from the image of the artificial tissue (or image segments extracted from the image) to determine a QC report which provides an indication of the maturation state of the artificial tissue. This is shown in Figure 11.

[0161] Figure 11 shows a portion of a QC report obtained over the maturation period of the artificial tissue shown in Figure 10 according to an embodiment of the present disclosure.

[0162] Figure 11 shows a first chart 1102, a second chart 1104, and a third chart 1106. In one embodiment, the first chart 1102, the second chart 1104, and the third chart 1106 are included as part of a QC report which records the maturation state of an artificial tissue during a maturation period. The QC report shown in Figure 11 is generated at a predetermined time point corresponding to 49 days post-seeding and comprises the tissue parameters generated at the preceding predetermined time points (e.g., 1 day postseeding, 2 days post-seeding, etc.). The QC report shown in Figure 11 provides important feedback and metrics related to the state of the artificial tissue at the predetermined time point (49 days post-seeding) as well as the development of the artificial tissue over the maturation period.

[0163] The first chart 1102 comprises a plot of tissue length obtained during maturation of an artificial tissue. The first chart 1102 comprises three points corresponding to the length of the artificial tissue obtained 7 days post-seeding, 14 days post-seeding, and 49 days post-seeding. The lengths plotted in the first chart 1102 are obtained from the third image 1006, the fourth image 1008, and the fifth image 1010 shown in Figure 10 using a segmentation unit (i.e., the segmentation unit 104 in Figure 1) and a feature extraction unit (i.e., the extraction unit 106 in Figure 1) as described in more detail above in relation to Figure 5.

[0164] The second chart 1104 comprises a plot of tissue widths obtained during maturation of the artificial tissue. The second chart 1104 comprises five plots of tissue width (obtained at the plurality of mid-points 512-1, 512-2, 512-3, 512-4 and the centroid 506 as shown in Figure 5) obtained 7 days post-seeding, 14 days post-seeding, and 49 days post-seeding. The widths plotted in the second chart 1104 are obtained from the third image 1006, the fourth image 1008, and the fifth image 1010 shown in Figure 10 using a segmentation unit (i.e., the segmentation unit 104 in Figure 1) and a feature extraction unit (i.e., the extraction unit 106 in Figure 1) as described in more detail above in relation to Figure 5. The plot marked "Wl" corresponds to the tissue width measured at the first mid-point 512- 1 of the plurality of mid-points shown in Figure 5. The plot marked "W2" corresponds to the tissue width measured at the second mid-point 512-2 of the plurality of mid-points shown in Figure 5. The plot marked "W3" corresponds to the tissue width measured at the centroid 506 shown in Figure 5. The plot marked "W4" corresponds to the tissue width measured at the third mid-point 512-3 of the plurality of mid-points shown in Figure 5. The plot marked "W5" corresponds to the tissue width measured at the fourth midpoint 512-4 of the plurality of mid-points shown in Figure 5.

[0165] The third chart 1106 comprises a plot of passive wire tension (or passive wire curvature, PWC) obtained during maturation of the artificial tissue. The third chart 1106 comprises two plots of passive wire tension obtained 1 day post-seeding, 2 days post-seeding, 7 days post-seeding, 14 days post-seeding, and 49 days post-seeding. The passive wire tension plotted in the third chart 1106 are obtained from the first image 1002, the second image 1004, the third image 1006, the fourth image 1008, and the fifth image 1010 shown in Figure 10 using a segmentation unit (i.e., the segmentation unit 104 in Figure 1) and a feature extraction unit (i.e., the extraction unit 106 in Figure 1) as described in more detail above in relation to Figure 5.

[0166] Figure 12 shows a tissue attachment grading portion of a QC report obtained over the maturation period of the artificial tissue shown in Figure 10 according to an embodiment of the present disclosure.

[0167] Figure 12 shows a first chart 1202, a second chart 1204, a third chart 1206, and a fourth chart 1208. In one embodiment, the first chart 1202, the second chart 1204, and the third chart 1206 are included as part of a QC report which records the maturation state of an artificial tissue during a maturation period. In an embodiment, the tissue grading portion of the QC report shown in Figure 12 is a part of the QC report shown in Figure 11. The tissue grading portion of the QC report shown in Figure 12 is generated at a plurality of predetermined time points (as discussed above in relation to Figure 10) and provides important feedback and metrics related to the state of the artificial tissue as encoded by the tissue grading.

[0168] The first chart 1202 comprises a plot of the tissue attachment grading of the top left portion of the artificial tissue in relation to the side of the well near to the tissue scaffold. At 7 days post-seeding (as shown in the third image 1006 of Figure 10) there is no attachment of the top left portion of the artificial tissue to the attachment site (i.e., the side of the well near to the tissue scaffold) and the artificial tissue is translucent, as indicated by the black filled circle at the 7 days post-seeding time point. At 14 days post seeding (as shown in the fourth image 1008 of Figure 10), the top left portion of the artificial tissue is not translucent, as indicated by the white circle at the 14 days postseeding time point, and has a major, or severe, attachment to the attachment site. The tissue attachment grading at 49 days post-seeding (as shown in the fifth image 1010 of Figure 10) for the top left portion of the artificial tissue is the same as the attachment grading at 14 days post-seeding.

[0169] The second chart 1204 comprises a plot of the tissue attachment grading of the top right portion of the artificial tissue in relation to the side of the well near to the tissue scaffold. At all predetermined time points— 7 days post-seeding, 14 days post-seeding, and 49 days post-seeding— there is no attachment of the top right portion of the artificial tissue to the attachment site (the side of the well near to the tissue scaffold).

[0170] The third chart 1206 comprises a plot of the tissue attachment grading of the bottom left portion of the artificial tissue in relation to the side of the well near to the tissue scaffold. At 7 days post-seeding (as shown in the third image 1006 of Figure 10) there is no attachment of the bottom left portion of the artificial tissue to the attachment site and the artificial tissue is translucent. At 14 days post seeding (as shown in the fourth image 1008 of Figure 10), the bottom left portion of the artificial tissue is translucent and has a minor attachment to the attachment site. At 49 days post-seeding (as shown in the fifth image 1010 of Figure 10), the bottom left portion of the artificial tissue is not translucent and has a major, or severe, attachment to the attachment site.

[0171] The fourth chart 1208 comprises a plot of the tissue attachment grading of the bottom right portion of the artificial tissue in relation to the side of the well near to the tissue scaffold. At all predetermined time points— 7 days post-seeding, 14 days post-seeding, and 49 days post-seeding— there is no attachment of the bottom right portion of the artificial tissue to the attachment site (the side of the well near to the tissue scaffold).

[0172] Using the information provided in the portion of the QC report shown in Figure 12, it is possible to determine that the artificial tissue has an irregular attachment which is limited to the left hand side of the artificial tissue only. This information can be used to modify of audit the manufacturing process and / or provide an explanation as to variability in experimental outputs.

[0173] Whilst the above description is directed to the use of QC reports during the maturation process of an artificial tissue, tissue parameters, tissue descriptors, and / or QC reports can be used post-maturation. For example, QC reports can be generated at time points when experimental data is generated or extracted from an artificial tissue. In such examples, the QC reports (and / or the tissue parameters and tissue descriptors) provide an indication of the state of the artificial tissue at the time points when the data is generated or extracted. As such, the QC reports can be used to gain explainable insights into experimental outcomes and / or track developmental, physiological, or morphological change in an artificial tissue during an experimental procedure.

[0174] Having described the functionality of the quality control (QC) unit 108 shown in Figure 1, the description will now turn to the assay unit 110 shown in Figure 1.

[0175] In general, the assay unit 110 determines a change in artificial tissue phenotype, as encoded by a tissue descriptor generated by the extraction unit 106. More particularly, the assay unit 110 is configured to compare a first tissue descriptor which encodes a phenotype of an artificial tissue under a first set of conditions to a second tissue descriptor which encodes a phenotype of the artificial tissue under a second set of conditions. The first and second set of conditions can be reference conditions (e.g., control conditions) or perturbed conditions associated with a perturbation of the artificial tissue such as a drug treatment, a disease model, a different cell line, or a physical perturbation of the artificial tissue. For example, the first set of conditions may be control conditions and the second set of conditions may be perturbed conditions associated with a drug treatment. The relative change in tissue descriptors provides a comparable encoding or featurization of the change in phenotype as a result of the change from the first set of conditions to the second set of conditions. The relative change can be used as a feature, or feature vector, in several downstream tasks such as drug development and discovery.

[0176] For example, a first tissue descriptor is obtained for an artificial tissue under reference conditions. The reference conditions comprise a vehicle treatment of the artificial tissue. The first tissue descriptor comprises an estimated smoothness of the artificial tissue obtained using the segmentation unit 104 and extraction unit 106 of the system 100 of Figure 1 as described in detail above. Because the first tissue descriptor is associated with the artificial tissue under reference conditions, it is considered an encoding of a baseline phenotype of the artificial tissue. A second tissue descriptor is obtained for the artificial tissue under perturbed conditions. The perturbed conditions comprise a drug treatment of the artificial tissue according to a first dosage of a compound. That is, at a time point between obtaining the first tissue descriptor and obtaining the second tissue descriptor, a first dosage of the compound is applied to the artificial tissue. The second tissue descriptor comprises an estimated smoothness of the artificial tissue. The second tissue descriptor is considered an encoding of a phenotype of the artificial tissue under perturbed conditions as a result of the first dosage of the compound. The assay unit 110 is configured to compare the second tissue descriptor to the first tissue descriptor to determine a phenotypic change in the artificial tissue. That is, the relative change between the tissue descriptor obtained under reference conditions and the tissue descriptor obtained under perturbed conditions corresponds to a quantification of the phenotypic change of the artificial tissue occurring as a result of the change in conditions. The relative change is then compared to a database of predetermined relative changes obtained from artificial tissues under conditions with a known effect (e.g., a change from reference conditions to a perturbed set of conditions associated with a drug having a known mechanism of action). Based on the comparison, one or more closely matching relative changes are identified from the database. Because the relative changes identified are similar to those exhibited by the change of conditions, it can be determined that the compound applied to the artificial tissue may have an effect(s) similar to those associated with the closely matching relative changes (e.g., a similar toxicity, mechanism of action, or the like). Thus, the geometric tissue descriptors described in the present disclosure provide an efficient and discriminative featurization of the phenotype of an artificial tissue which can be used to identify one or more effects associated with unknown compounds.

[0177] Figure 13 illustrates a change in tissue phenotype as encoded by a geometric tissue descriptor according to an embodiment of the present disclosure.

[0178] Figure 13 shows a plot 1300 of tissue volume for artificial tissues treated by dimethyl sulfoxide (DMSO), treated by 2 .m. of a pharmaceutical research compound hypothesized to have an effect on tissue volume (referred to in relation to Figure 13 as "the compound"), and treated by 10 / zm of the compound. The DMSO plot comprises tissue volume measurements obtained from a group of eight engineered cardiac tissue samples under control conditions. The DMSO group have a mean tissue volume of 0.3374mm3with a standard deviation of 0.0214mm3. The first treatment plot (2 / zm of the compound) comprises tissue volume measurements obtained from a group of five engineered cardiac tissue samples under perturbed conditions corresponding to a 2 / zm dose of the compound. The first treatment group have a mean tissue volume of 0.4125mm3with a standard deviation of 0.0226mm3. The second treatment plot (10 / zm of the compound) comprises tissue volume measurements obtained from a group of seven engineered cardiac tissue samples under perturbed conditions corresponding to a 10 / zm dose of the compound. The second treatment group have a mean tissue volume of 0.4433mm3with a standard deviation of 0.0305mm3.

[0179] The relative difference between the DMSO group and the first treatment group is shown as Si in Figure 13. The relative difference between the DMSO group and the second treatment group is shown as S2in Figure 13. The relative difference between the first treatment group and the second treatment group is shown as S3in Figure 13. Each of the relative differences shown in Figure 13 can be used as an encoding of phenotypic change. For example, the relative difference encodes the phenotypic change from reference conditions to perturbed conditions comprising a treatment of the compound at 2 .m.

[0180] An analysis of variance (ANOVA) test applied to the three groups shows significant differences in volume between the groups (p < 9.6e - 7). A Tukey test showed the DMSO group volumes significantly differed from the first treatment group (p < 1.9e - 4) and the second treatment group (p < 8.7e - 7) with no statistically significant difference between the first and the second treatment groups (p = 0.12).

[0181] Therefore, the geometric encoding of tissue phenotype provided by the tissue volume descriptor provides a powerful and discriminative feature which can be efficiently used for a number of downstream drug discovery and development tasks.

[0182] The description will now turn to methods for utilizing components, processes, and operations of the above described systems to perform quality control of artificial tissue, tissue attachment grading, and geometric encoding of tissue phenotype.

[0183] Figure 14 shows a method 1400 for quality control (QC) of artificial tissue according to an aspect of the present disclosure.

[0184] The method 1400 comprises the steps of obtaining 1402 an image, extracting 1404 one or more image segments, determining 1406 one or more tissue parameters, and outputting 1408 a QC report. In one embodiment, the method 1400 is performed by the system 100 shown in Figure 1.

[0185] In general, the QC report generated by the method 1400 provides an automatic indication of the state of an artificial tissue at a predetermined time point as encoded within one or more tissue parameters. The automation of quality control, particularly during the maturation period of artificial tissue, helps to decrease the requirement for human involvement thereby decreasing user assessment bias and increasing the scalability of tissue engineering systems. Moreover, the QC reports generated by the method 1400 can supplement existing tissue assays to improve the explainability of such models in the presence of experimental variability (e.g., by providing quantitative explanations of the physiological state of the artificial tissue).

[0186] At the step of obtaining 1402, a first image of an artificial tissue is obtained at a first predetermined time point. The first image comprises an image of the artificial tissue taken at the first predetermined time point. The artificial tissue comprises a muscle tissue such as a cardiac muscle tissue or a skeletal muscle tissue. In one embodiment, the first image comprises a brightfield microscopy image obtained from a sensor assembly of the bioreactor.

[0187] In an embodiment, the first predetermined time point comprises a predetermined time point during a maturation period of the artificial tissue. For example, the predetermined time point is either 1 day post-seeding, 2 days post-seeding, 7 days post-seeding, 14 days post-seeding, or 49 days post-seeding. The first image can be automatically obtained (e.g., by a control unit or system such as the system 100 shown in Figure 1) at the predetermined time point.

[0188] At the step of extracting 1404, one or more image segments are extracted from the first image. Each of the one or more image segments comprise a region of interest of the artificial tissue within the first image (e.g., the well region 202, the first tissue scaffold region 204, the second tissue scaffold region 206 and the tissue region 208 shown in Figure 2).

[0189] As described in more detail above in relation to Figure 5, the one or more image segments are extracted from the first image using a segmentation unit (e.g., the segmentation unit 104 of the system 100 in Figure 1). In one embodiment, the one or more image segments are extracted from the first image using a segmentation pipeline. The segmentation pipeline utilizes a two-step segmentation process. The first segmentation step segments the well region (i.e., the well region 202 shown in Figure 2) from the first image. The second segmentation step segments the tissue region (i.e., the tissue region 208 shown in Figure 2) from the first image cropped to the well region identified in the first segmentation step. In an embodiment, the segmentation pipeline comprises a plurality of sequential morphological operations including one or more of an image equalization operation, an affine transformation, an image crop operation, and a binarization (as described above in relation to the segmentation pipeline shown in Figures 3 and 4).

[0190] At the step of determining 1406, one or more tissue parameters are determined from the one or more image segments. The one or more tissue parameters are indicative of a physiological state of the artificial tissue at the first predetermined time point. The physiological state of the artificial tissue includes one or more of: a maturation state, a presence of disease phenotype, a drug treatment effect, and an irregularity associated with the tissue.

[0191] In one embodiment, the one or more tissue parameters comprise one or more of a tissue irregularity grading; a geometric parameter; and a passive tension measure. The tissue irregularity grading is a non-geometric feature which quantifies irregularities, or problems, related to the artificial tissue including holes or discontinuities within the artificial tissue, irregularities in tissue shape, debris within the tissue, a broken tissue, and a broken tissue scaffold. In one embodiment, the tissue irregularity grading comprises a tissue attachment grading associated with attachments of the artificial tissue to the device within which the artificial tissue is contained. The tissue attachment grading is indicative of a degree of attachment of a portion of the artificial tissue to an attachment site of the device (e.g., a tissue scaffold of the device, a side wall of the device near to the tissue scaffold, or a side wall of the device away from the tissue scaffold). The tissue attachment grading comprises a plurality of scores including an attachment score (i.e., a score or label for the attachment output level 704 shown in Figure 7), an attachment severity score (i.e., a score or label for the severity output level 706 shown in Figure 7), and a translucence score (i.e., a score or label for the translucence output level 708 shown in Figure 7). The tissue attachment grading is determined using a grading model based on an image segment of the one or more image segments which includes the portion of the tissue (e.g., a quadrant image such as those shown in Figure 6). The grading model can be a hierarchy of classifiers, or a trained multi-task learning (MTL) network as shown in Figures 9A-9C and described in detail above. The trained MTL network comprises an input convolutional neural network configured to receive an input image (e.g., the image segment of the one or more image segments), a plurality of independent output networks each configured to estimate a score associated with the input image (e.g., an attachment score, an attachment severity score, a translucence score, and a grading score), and a shared network coupled between the input convolutional neural network and each of the plurality of independent output networks.

[0192] As described in more detail in relation to Figure 5 above, the geometric parameter comprises one of: a tissue length, a tissue width, a tissue volume, or a tissue area. The geometric parameter is determined from an image segment of the one or more image segments extracted at the step of extracting 1404. The image segment comprises a region of the first image which circumscribes the tissue (e.g., the tissue region 502 shown in Figure 5). The tissue length is associated with a length of a major axis of the image segment (e.g., the length between the first boundary point 508 and the second boundary point 510 of the tissue region 502 shown in Figure 5). The tissue width is associated with a length of a minor axis of the image segment at a predetermined relative position along the major axis of the tissue (e.g., a minor axis length at the first mid-point 512-1, the second mid-point 512-2, the centroid 506, etc.). The tissue volume is determined from a plurality of tissue widths according to a slice volume model. As described in more detail in relation to Figure 5 above, the slice volume model estimates a volume of a slice of the artificial tissue at a location along the longitudinal axis based on an estimated cross- sectional area of the artificial tissue determined from a width of the artificial tissue at the location. The estimated cross-sectional area is determined according to a model parameterized by the width of the artificial tissue and a predetermined tissue elongation parameter, as shown in Equation (1). The predetermined tissue elongation parameter is in the range of from 1 to 5. The tissue area the estimated surface area of the artificial tissue captured within the image.

[0193] The passive tension measure corresponds to the estimated tension that the artificial tissue exerts on a tissue scaffold when at rest (i.e., in the absence of any external stimulation). The passive tension measure is determined from an image segment of the one or more image segments which includes a tissue scaffold of the device (e.g., the first tissue scaffold region 204 shown in Figure 2). The passive tension measure is estimated from a model of displacement of the tissue scaffold. The model of displacement comprises one or more linear models or a linear model and a quadratic model. The model estimates the (vertical) distance between the anchor points of the tissue scaffold and the most extreme curvature point of the first tissue scaffold.

[0194] At the step of outputting 1408, a QC report is output based on the one or more tissue parameters.

[0195] The QC report provides a representation of the physiological state of an artificial tissue at the first predetermined time point and can track the physiological development and / or change in physiological state of the artificial tissue over the predetermined time points leading up to the first predetermined time point. The QC report is generated from the one or more tissue parameters such that the QC report provides an indication of the physiological state of the artificial tissue (and the change thereto) at the predetermined time point as encoded within the one or more tissue parameters. Example portions of a QC report are shown in Figures 10 and 12 as described in more detail above.

[0196] Additionally, or alternatively, the QC report comprises a tissue quality alert when at least one of the tissue parameters does not meet a predetermined quality control threshold. For example, a quality control threshold may require that the tissue attachment grading scores for each of the tissue's attachment points (i.e., the four corners of the artificial tissue) are to be above a threshold value at a given time point (e.g., 49 days post-seeding). In one embodiment, method 1400 further comprises the step of causing a command to be transmitted to the bioreactor based on the tissue quality alert. The bioreactor may be configured to process the command (e.g., by shutting down the device containing the artificial tissue which fails to meet the quality control threshold, or issuing an alert at the bioreactor highlighting that there is a quality issue associated with the artificial tissue which fails to meet the quality control threshold). Additionally, or alternatively, the command is transmitted to another device (e.g., a computing device, a display device, or the like) to alert a user of the system that the quality control threshold has or has not been met.

[0197] In one embodiment, outputting the QC report comprises storing, or saving, QC report to a persistent storage such as a non-volatile memory, a non-transitory medium, or the like. Additionally, or alternatively, outputting the QC report comprises transmitting the QC report via a network (e.g., a local area network, a wide area network, and the like), or displaying the QC report, or a portion thereof, for review by a user.

[0198] Figure 15 shows a method 1500 for engineered tissue attachment grading according to an aspect of the present disclosure.

[0199] The method 1500 comprises the step of obtaining 1502 an image, extracting 1504 a first region, determining 1506 a plurality of attachment scores, determining 1508 an attachment grading, and outputting 1510 the attachment grading. In one embodiment, the steps of extracting 1504, determining 1506, determining 1508, and outputting 1510 are repeated for multiple regions extracted from the image. For example, the steps are repeated four times such that attachment gradings are determined for each quadrant image extracted from the image obtained at the step of obtaining 1502. In one embodiment, the method 1500 is performed by the system 100 shown in Figure 1.

[0200] In general, the method 1500 uses a multi-task learning (MTL) network to determine an attachment grading for an artificial tissue at a respective attachment site. Attachments of the artificial tissue to elements of the device which are not a tissue scaffold, or inadequate attachment of the artificial tissue to the tissue scaffold, may inhibit the functional response of the artificial tissue thus causing irregular or inaccurate phenotype measurements or functional response values. As such, tissue attachment gradings determined by the method 1500 help identify and quantify issues regarding the state of the artificial tissue (e.g., during maturation) which helps improve the quality of artificial tissue produced by identifying sources of systemic manufacturing variability. During an example QC process, artificial tissues identified as having tissue attachment gradings which meet certain criteria (e.g., a severe attachment to the side of the well away from the tissue scaffold) can be removed thereby allowing new artificial tissue to be grown in place.

[0201] At the step of obtaining 1502, an image of an engineered tissue grown within a device is obtained. The device comprises a tissue scaffold for attachment to the engineered tissue. The image is obtained either directly, or indirectly, from a bioreactor within which the artificial tissue is contained (e.g., the bioreactor 102 shown in Figure 1). The bioreactor comprises a device including the tissue scaffold. The artificial tissue comprises engineered muscle tissue such as cardiac tissue or skeletal muscle tissue. The first set of conditions associated with the artificial tissue comprise reference or control conditions such as a vehicle treatment of the artificial tissue. Alternatively, the first set of conditions comprise perturbed conditions associated with a perturbation of the artificial tissue such as a drug treatment, a disease model, a different cell line, or a physical perturbation of the artificial tissue.

[0202] At the step of extracting 1504, a first region is extracted from the image. The first region of the image includes a first portion of the engineered tissue and a portion of the tissue scaffold.

[0203] As stated in more detail above in relation to Figures 5 and 6, the first region corresponds to a quadrant of the image which comprises a portion (i.e., "corner") of the engineered tissue and a portion of the tissue scaffold (as shown in Figure 6).

[0204] At the step of determining 1506, a plurality of attachment scores are determined using a multi-task learning (MTL) network based on the first region.

[0205] The architecture of the MTL network is described in more detail in relation to Figure 9 above. The MTL network comprises an input convolutional neural network configured to receive an input image (e.g., the input network 902 shown in Figure 9A), a plurality of independent output networks each configured to estimate an attachment score associated with the input image (e.g., the attachment network 906-1, the severity network 906-2, and the translucence network 906-3, the auxiliary network 906-4, and in some embodiments the minor severity network 906-5 shown in Figure 9A), and a shared network coupled between the input convolutional neural network and each of the plurality of independent output networks (e.g., the shared dense network 904 shown in Figure 9A). The input convolutional neural network comprise a plurality of convolutional neural networks including a plurality of inception layers. The shared network comprises at least one pooling layer and a plurality of dense layers. Each of the plurality of independent output networks comprise a plurality of dense layers and at least one output layer. In one embodiment, the plurality of independent output networks further comprise an auxiliary output network configured to predict an attachment grading (e.g., the auxiliary output network 906-4 shown in Figure 9A).

[0206] The plurality of attachment scores include an attachment indicator (i.e., a score or label for the attachment output level 704 shown in Figure 7), an attachment severity score (i.e., a score or label for the severity output level 706 shown in Figure 7), and a translucence score (i.e., a score or label for the translucence output level 708 shown in Figure 7).

[0207] In one embodiment, the method 1500 further comprises the step of training the MTL network. As described above in relation to Figure 9, the MTL network is trained on a plurality of images of engineered tissues and associated tissue scaffolds. Each of the plurality of images is associated with a corresponding plurality of attachment scores.

[0208] At the step of determining 1508, an attachment grading for the first region is determined based on the plurality of attachment scores. The attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to an attachment site of the device. In one embodiment, the attachment grading comprises a combination of the plurality of attachment scores.

[0209] The attachment site to which the attachment grading is associated is one or the tissue scaffold (e.g., the attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to the portion of the tissue scaffold shown in the first image), a first portion of the side of the well of the device near to the tissue scaffold (e.g., the attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to the first portion of the side of the well), or a second portion of the side of the well of the device away from the tissue scaffold (e.g., the attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to the second portion of the side of the well).

[0210] At the step of outputting 1510, the attachment grading is output.

[0211] In one embodiment, outputting the attachment grading comprises storing, or saving, the attachment grading to a persistent storage such as a non-volatile memory, a non- transitory medium, or the like. Additionally, or alternatively, outputting the attachment grading comprises transmitting the attachment grading via a network (e.g., a local area network, a wide area network, and the like), or displaying the attachment grading, or a portion thereof, for review by a user. Additionally, or alternatively, outputting the attachment grading comprises including the attachment grading, or a portion thereof, in a QC report (e.g., as shown in Figure 12 above).

[0212] Additionally, or alternatively, the attachment grading is output as part of an alert when the attachment grading does not meet a predetermined attachment criterion. In one embodiment, a command is caused to be transmitted to the bioreactor containing the first tissue based on the alert. The bioreactor may be configured to process the command (e.g., by shutting down the device containing the artificial tissue which fails to meet the predetermined attachment criterion, or issuing an alert at the bioreactor highlighting that there is a quality issue associated with the artificial tissue which fails to meet the predetermined attachment criterion). Additionally, or alternatively, the command is transmitted to another device (e.g., a computing device, a display device, or the like) to alert a user of the system that the predetermined attachment criterion has or has not been met.

[0213] Figure 16 shows a method 1600 for geometric encoding of tissue phenotype according to an aspect of the present disclosure.

[0214] The method 1600 comprises the steps of obtaining 1602 an image, extracting 1604 a first region from the image, extracting 1606 one or more geometric features from the first region, and generating 1608 a first tissue descriptor. The method 1600 further comprises the optional step of outputting 1610 the tissue descriptor. In one embodiment, the method 1600 is performed by the system 100 shown in Figure 1 and described in detail above.

[0215] In general, the method 1600 provides an efficient mechanism for encoding the phenotype of an artificial tissue from one or more geometric features extracted from an image of the artificial tissue. The geometric features provide a shape-based characterization of the artificial tissue which can be used as features to determine the phenotypic change in the artificial tissue.

[0216] At the step of obtaining 1602, an image of an artificial tissue under a first set of conditions is obtained (e.g., the image 116 shown in Figure 1). The image is obtained either directly, or indirectly, from a bioreactor within which the artificial tissue is contained (e.g., the bioreactor 102 shown in Figure 1). For example, the image comprises a brightfield microscopy image obtained from a sensor assembly of a bioreactor.

[0217] The artificial tissue comprises engineered muscle tissue such as cardiac tissue or skeletal muscle tissue. The first set of conditions associated with the artificial tissue comprise reference or control conditions such as a vehicle treatment of the artificial tissue. Alternatively, the first set of conditions comprise perturbed conditions associated with a perturbation of the artificial tissue such as a drug treatment, a disease model, a different cell line, or a physical perturbation of the artificial tissue.

[0218] At the step of extracting 1604, a first region is extracted from the image such that the first region circumscribes the artificial tissue within the image. The first region is extracted using a segmentation unit such as the segmentation unit 104 shown in Figure 1. The first region corresponds to a portion of the image containing a specific part of the artificial tissue and / or the device within which the artificial tissue is contained. More particularly, the first region corresponds to a tissue region (such as the tissue region 208 shown in Figure 2).

[0219] The first region is extracted from the image using a segmentation pipeline. The segmentation pipeline utilizes a two-step segmentation process. The first segmentation step segments the well region (i.e., the well region 202 shown in Figure 2) from the image. The second segmentation step segments the tissue region (i.e., the tissue region 208 shown in Figure 2) from the image cropped to the well region identified in the first segmentation step. In an embodiment, the segmentation pipeline comprises a plurality of sequential morphological operations including one or more of an image equalization operation, an affine transformation, an image crop operation, and a binarization (as described above in relation to the segmentation pipeline shown in Figures 3 and 4).

[0220] At the step of extracting 1606, one or more geometric features are extracted from the first region. The one or more geometric features provide a shape-based characterization of the artificial tissue.

[0221] The one or more geometric features, or parameters, comprise a plurality of tissue widths determined at a corresponding plurality of locations along a longitudinal axis of the first region (e.g., locations corresponding to the plurality of mid-points 512-1, 512-2, 512-3, 512-4, and the centroid 506 shown in Figure 5). Each of the plurality of tissue widths is perpendicular to the longitudinal axis of the first region. The corresponding locations are evenly spaced along the longitudinal axis of the first region.

[0222] Additionally, or alternatively, the one or more geometric features, or parameters, comprise a tissue length corresponding to the longitudinal length of the artificial tissue.

[0223] Additionally, or alternatively, the one or more geometric features, or parameters, comprise a tissue area corresponding to the estimated surface area of the artificial tissue captured within the image.

[0224] Additionally, or alternatively, the one or more geometric features, or parameters, comprise a tissue volume corresponding to an estimated volume of the artificial tissue. The tissue volume is determined from the plurality of tissue widths according to a slice volume model. As described in more detail in relation to Figure 5 above, the slice volume model estimates a volume of a slice of the artificial tissue at a location along the longitudinal axis based on an estimated cross-sectional area of the artificial tissue determined from a width of the artificial tissue at the location. The estimated cross-sectional area is determined according to a model parameterized by the width of the artificial tissue and a predetermined tissue elongation parameter, as shown in Equation (1). The predetermined tissue elongation parameter is in the range of from 1 to 5.

[0225] At the step of generating 1608, a first tissue descriptor is generated based on the one or more geometric features. The first tissue descriptor encodes a phenotype of the artificial tissue under the first set of conditions.

[0226] In one embodiment, the first tissue descriptor comprises an estimated smoothness of the artificial tissue determined from a variance of the plurality of tissue widths, as described in detail in relation to Equation (2) above. In another embodiment, the first tissue descriptor comprises an estimated volume of the artificial tissue. That is, the first tissue descriptor corresponds to the tissue volume geometric feature. In another embodiment, the first tissue descriptor comprises an estimated area of the artificial tissue. That is, the first tissue descriptor corresponds to the estimated area geometric feature.

[0227] At the optional step of outputting 1610, the first tissue descriptor is output. In one embodiment, outputting the first tissue descriptor comprises storing, or saving, the first tissue descriptor to a persistent storage such as a non-volatile memory, a non-transitory medium, or the like. Additionally, or alternatively, outputting the first tissue descriptor comprises transmitting the first tissue descriptor via a network (e.g., a local area network, a wide area network, and the like), or displaying the first tissue descriptor for review by a user.

[0228] Figure 17 shows a method 1700 for determining an effect associated with a phenotypic change according to an embodiment of the present disclosure.

[0229] The method 1700 comprises the steps of obtaining 1702 a first tissue descriptor, obtaining 1704 a second tissue descriptor, comparing 1706 the tissue descriptors, determining 1708 a phenotypic change, comparing 1710 the phenotypic change to one or more predetermined changes, and determining 1712 an effect. The method 1700 further comprises the optional step of outputting 1714 the effect. In one embodiment, the method 1700 is performed by the system 100 shown in Figure 1 and described in detail above.

[0230] In general, the method 1700 corresponds to a technique for determining an effect associated with a phenotypic change in an artificial tissue from a first set of condition (e.g., reference conditions) to a second set of conditions (e.g., perturbed conditions). The method 1700 utilizes a relative change in (geometric) tissue descriptors between the two conditions thereby providing a comparable encoding, or featurization, of the change in phenotype as a result of the change from the first set of conditions to the second set of conditions. By mapping the relative change in tissue descriptors to a database of predetermined tissue descriptors, an effect associated with the change from the first set of conditions to the second set of conditions can be efficiently determined.

[0231] At the step of obtaining 1702, a first tissue descriptor is obtained. The first tissue descriptor is associated with an artificial tissue under a first set of conditions. In one embodiment, the first tissue descriptor is obtained using the method 1600 shown in Figure 16. Alternatively, the first tissue descriptor is obtained from a persistent storage device or storage medium, or via a network connection or other transmission medium.

[0232] At the step of obtaining 1704, a second tissue descriptor is obtained. The second tissue descriptor is associated with the artificial tissue (to which the first tissue descriptor relates) under a second set of conditions. In one embodiment, the second tissue descriptor is obtained using the method 1600 shown in Figure 16. Alternatively, the second tissue descriptor is obtained from a persistent storage device or storage medium, or via a network connection or other transmission medium.

[0233] In one embodiment, the first set of conditions (associated with the first tissue descriptor) are reference, or control, conditions and the second set of conditions are perturbed conditions. The perturbed conditions are associated with a perturbation of the artificial tissue such as a drug treatment, a disease model, a different cell line, or a physical perturbation of the artificial tissue.

[0234] The first tissue descriptor and the second tissue descriptor encode a phenotype of the artificial tissue under the different conditions according to the same set of geometric features. For example, both the first tissue descriptor and the second tissue descriptor can be an estimated smoothness of the artificial tissue, an estimate volume of the artificial tissue, or the like.

[0235] At the step of comparing 1706, the first tissue descriptor is compared to the second tissue descriptor and at the subsequent step of determining 1708, a phenotypic change in the artificial tissue is determined based on the comparison performed at the step of comparing 1706.

[0236] Because the first tissue descriptor and the second tissue descriptor are quantitative, numerical, representations of the phenotype of the artificial tissue under different conditions, a comparison of the two tissue descriptors provides a quantitative, numerical, representation of the phenotypic change in the artificial tissue from the first set of conditions to the second set of conditions. This is illustrated by the relative changes 6^62,63 between the conditions shown in Figure 13.

[0237] At the step of comparing 1710, the phenotypic change is compared to one or more predetermined phenotypic changes. Each of the one or more predetermined phenotypic changes are associated with a corresponding effect.

[0238] The one or more predetermined phenotypic changes are persistently stored in a database such as a relational or graph-based database, a static file, or another structured representation. Each of the one or more predetermined phenotypic changes comprises a relative change and at least one associated effect. For example, a predetermined phenotypic change may correspond to the relative change in tissue volume for an artificial cardiac tissue under control conditions compared to the artificial cardiac tissue under a first dose of a compound having a known toxicity; the effect associated with the relative change would thus correspond to the known toxicity of the compound. As a further example, a predetermined phenotypic change may correspond to the relative change in tissue area for an artificial cardiac tissue under conditions corresponding to a certain disease state compared to the artificial cardiac tissue having been subsequently dosed with a first drug treatment having a known mechanism of action; the effect associated with the relative change would thus correspond to the known mechanism of action of the first drug treatment.

[0239] At the step of determining 1712, an effect associated with the second set of conditions is determined based on the comparison performed at the step of comparing 1710.

[0240] The step of comparing 1710 comprises searching through the one or more predetermined phenotypic changes and identifying a similarity, or dissimilarity, between the phenotypic change and the one or more predetermined phenotypic change. The one or more predetermined phenotypic changes are searched using either a linear or sequential search or a binary search. As a result of the step of comparing 1710, one or more candidate predetermined phenotypic changes are identified (i.e., the predetermined phenotypic changes which are most similar to the phenotypic change). Based on the one or more candidate predetermined phenotypic changes the effect associated with the second set of conditions is determined. For example, if the dissimilarity between the most similar predetermined phenotypic change and the phenotypic change is less than a predetermined threshold, then the effect associated with the most similar predetermined phenotypic change is associated with the second set of conditions.

[0241] At the optional step of outputting 1714, the effect is output. In one embodiment, outputting the effect comprises storing, or saving, the effect to a persistent storage such as a non-volatile memory, a non-transitory medium, or the like. Additionally, or alternatively, outputting the effect comprises transmitting the effect via a network (e.g., a local area network, a wide area network, and the like), or displaying the effect for review by a user.

[0242] Figure 18 shows a bioreactor system 1800 according to embodiments of the present disclosure.

[0243] The bioreactor system 1800 comprises a bioreactor 1802 and a control unit 1804. The bioreactor 1802 comprises a device 1806 for growing engineered tissues, a sensor assembly 1808, and an interface 1810. The interface 1810 communicatively couples the bioreactor 1802 and the control unit 1804 such that data may be exchanged between the bioreactor 1802 and the control unit 1804. In one embodiment, the bioreactor 1802 corresponds to the bioreactor 102 of the system 100 shown in Figure 1 and the control unit 1804 corresponds to a subsystem of the system 100 such as the first subsystem 122 or the second subsystem 124.

[0244] As shown in the expanded portion 1806-1 of the device 1806, the device 1806, or substrate, comprises one or more wells, such as the well 1814, one or more cell culture wells, such as the cell culture well 1816, a pair of electrodes including a first electrode 1818-1 and a second electrode 1818-2, and a pair of elements including a first element 1820-1 and a second element 1820-2. The well 1814 is positioned within the cell culture well 1816 and has a bottom on the device 1806, a first end 1822-1, and a second end 1822-2. The well 1814 is configured for growing an engineered tissue 1824 from cells seeded therein. Culture medium may be added to the cell culture well 1816 for growing and / or sustaining the engineered tissue 1824. The engineered tissue 1824, alternatively referred to as artificial tissue, comprises engineered muscle tissue. In one embodiment, the engineered muscle tissue is engineered cardiac tissue. In an alternative embodiment, the engineered muscle tissue is engineered skeletal muscle tissue.

[0245] The pair of electrodes are separated by a gap within which the well 1814 is positioned. The pair of electrodes are configured to apply an electrical stimulation to cell cultures within the one or more wells of the device 1806 (e.g., the engineered tissue 1824 within the well 1814 shown in the expanded portion 1806-1). During maturation of the cell cultures within the device 1806, the pair of electrodes apply stimulation to the cell cultures according to a multi-week electrical stimulation protocol. After the cell cultures are matured, the pair of electrodes may be configured to stimulate the cell cultures (e.g., the engineered tissue 1824) at a set frequency, or pacing frequency. The first element 1820-1 and the second element 1820-2 are disposed across the well 1814 such that there is a gap between the bottom of the well 1814 and the pair of elements. The first element 1820-1 and the second element 1820-2 are configured to: (a) permit attachment of the engineered tissue 1824 formed therebetween, thereby suspending the engineered tissue 1824 above the bottom of the well 1814, and (b) deform in response to the contractile force exerted on the pair of elements by the engineered tissue 1824, thereby simulating a physiological environment that is native to the engineered tissue 1824 and / or permitting measurement of the contractile force (e.g., by the sensor assembly 1808). For example, the pair of electrodes may subject the engineered tissue 1824 to an electrical stimulation at a frequency of 0.1Hz. The engineered tissue 1824 will contract in response to this electrical stimulation causing deformation of at least one of the first element 1820-1 and the second element 1820-2.

[0246] The sensor assembly 1808 is configured to obtain a plurality of images of a tissue (e.g., the engineered tissue 1824) at a plurality of time points. For example, the optical sensor may be configured to capture an image, or frame, of a tissue at a predetermined time point (e.g., 1 day post-seeding, 2 days post-seeding, etc.). In one embodiment, the control unit 1804 is configured to send a command 1826 to the bioreactor 1802 which causes the sensor assembly 1808 to obtain an image 1828 of a tissue (e.g., the engineered tissue 1824) which is then returned to the control unit 1804.

[0247] Figure 19 shows an example computing system for carrying out the methods of the present disclosure. Specifically, Figure 19 shows a block diagram of an embodiment of a computing system according to example embodiments of the present disclosure. The computing system shown in Figure 19 may correspond to a part, or the whole, of any of the functional units described above.

[0248] Computing system 1900 can be configured to perform any of the operations disclosed herein such as, for example, any of the operations discussed with reference to the functional units described in relation to Figure 1. Computing system includes one or more computing device(s) 1902. The one or more computing device(s) 1902 of computing system 1900 comprise one or more processors 1904 and memory 1906. One or more processors 1904 can be any general purpose processor(s) configured to execute a set of instructions, such as computing instructions including implemented in any one or more programming languages such as Python, Go, C, C+ + , C#, Java, or the like. For example, one or more processors 1904 can be one or more general-purpose processors, one or more field programmable gate array (FPGA), and / or one or more application specific integrated circuits (ASIC). In one embodiment, one or more processors 1904 include one processor. Alternatively, one or more processors 1904 include a plurality of processors that are operatively connected. One or more processors 1904 are communicatively coupled to memory 1906 via address bus 1908, control bus 1910, and data bus 1912. Memory 1906 can be a random access memory (RAM), a read only memory (ROM), a persistent storage device such as a hard drive, an erasable programmable read only memory (EPROM), and / or the like. The one or more computing device(s) 1902 further comprise I / O interface 1914 communicatively coupled to address bus 1908, control bus 1910, and data bus 1912.

[0249] Memory 1906 can store information that can be accessed by one or more processors 1904. For instance, memory 1906 (e.g., one or more non-transitory computer-readable storage mediums, memory devices) can include computer-readable instructions (not shown) (e.g., computing instructions) that can be executed by one or more processors 1904. The computer-readable instructions can be software written in any suitable programming language or can be implemented in hardware. Additionally, or alternatively, the computer- readable instructions can be executed in logically and / or virtually separate threads on one or more processors 1904. For example, memory 1906 can store instructions (not shown), such as computing instructions that when executed by one or more processors 1904 cause one or more processors 1904 to perform operations such as any of the operations and functions for which computing system 1900 is configured, as described herein. In addition, or alternatively, memory 1906 can store data (not shown) that can be obtained, received, accessed, written, manipulated, created, and / or stored. The data can include, for instance, the data and / or information described herein in relation to Figures 1 to 14. In some implementations, the one or more computing device(s) 1902 can obtain from and / or store data in one or more memory device(s) that are remote from the computing system 1900.

[0250] Computing system 1900 further comprises storage unit 1916, network interface 1918, input controller 1920, and output controller 1922. Storage unit 1916, network interface 1918, input controller 1920, and output controller 1922 are communicatively coupled to the central control unit (i.e., the memory 1906, the address bus 1908, the control bus 1910, and the data bus 1912) via I / O interface 1914.

[0251] Storage unit 1916 is a computer readable medium, preferably a non-transitory computer readable medium, comprising one or more programs, the one or more programs comprising computing instructions which when executed by the one or more processors 1904 cause computing system 1900 to perform the method steps of the present disclosure. Alternatively, storage unit 1916 is a transitory computer readable medium. Storage unit 1916 can be a persistent storage device such as a hard drive, a cloud storage device, or any other appropriate storage device. Network interface 1918 can be a Wi-Fi module, a network interface card, a Bluetooth module, and / or any other suitable wired or wireless communication device. In an embodiment, network interface 1918 is configured to connect to a network such as a local area network (LAN), or a wide area network (WAN), the Internet, or an intranet.

Claims

CLAIMSWhat is claimed is:

1. A method for geometric encoding of tissue phenotype, the method comprising: obtaining an image of an artificial tissue under a first set of conditions; extracting a first region from the image such that the first region circumscribes the artificial tissue within the image; extracting one or more geometric features from the first region, wherein the one or more geometric features provide a shape-based characterization of the artificial tissue; and generating a first tissue descriptor based on the one or more geometric features, wherein the first tissue descriptor encodes a phenotype of the artificial tissue under the first set of conditions.

2. The method of claim 1 further comprising: outputting the first tissue descriptor.

3. The method of claim 1 wherein the one or more geometric features comprise a plurality of tissue widths determined at a corresponding plurality of locations along a longitudinal axis of the first region, wherein each of the plurality of tissue widths is perpendicular to the longitudinal axis of the first region.

4. The method of claim 3 wherein the corresponding plurality of locations are evenly spaced along the longitudinal axis of the first region.

5. The method of claim 3 wherein the first tissue descriptor comprises an estimated smoothness of the artificial tissue determined from a variance of the plurality of tissue widths.

6. The method of claim 3 wherein the first tissue descriptor comprises an estimated volume of the artificial tissue determined from the plurality of tissue widths according to a slice volume model.

7. The method of claim 6 wherein the slice volume model estimates a volume of a slice of the artificial tissue at a location along the longitudinal axis based on an estimated cross-sectional area of the artificial tissue, wherein the estimated cross- sectional area is determined from a width of the artificial tissue at the location.

8. The method of claim 7 wherein the estimated cross-sectional area is determined according to a model parametrized by the width of the artificial tissue and a predetermined tissue elongation parameter.

9. The method of claim 8 wherein the predetermined tissue elongation parameter is in a range of from 1 to 5.

10. The method of claim 1 wherein the first tissue descriptor comprises an estimated area of the first region.

11. The method of claim 1 further comprising: obtaining a second tissue descriptor associated with the artificial tissue under a second set of conditions; and determining a phenotypic change in the artificial tissue based on a comparison of the first tissue descriptor and the second tissue descriptor.

12. The method of claim 11 further comprising: comparing the first tissue descriptor and the second tissue descriptor.

13. The method of claim 11 wherein the first set of conditions are reference conditions.

14. The method of claim 11 wherein the second set of conditions are perturbed conditions associated with a perturbation of the artificial tissue.

15. The method of claim 14 wherein the perturbation comprises one or more of: a drug treatment; a disease model; or a different cell line.

16. The method of claim 11 further comprising: comparing the phenotypic change to one or more predetermined phenotypic changes each of which associated with a corresponding effect; anddetermining an effect associated with the second set of conditions based on the comparing.

17. The method of claim 16 further comprising: outputting the effect.

18. The method of claim 16 wherein the effect comprises one or more of a mechanism of action or a toxicity.

19. The method of claim 1 wherein the first region is extracted from the image using a segmentation pipeline.

20. The method of claim 19 wherein the segmentation pipeline includes a plurality of sequential morphological operations.

21. The method of claim 19 wherein the segmentation pipeline includes one or more of: an image equalization operation; an affine transformation; an image crop operation; and a binarization.

22. The method of claim 1 wherein the image comprises a brightfield microscopy image.

23. The method of claim 1 wherein the image is obtained from a bioreactor within which the artificial tissue is grown.

24. The method of claim 1 wherein the artificial tissue comprises muscle tissue.

25. The method of claim 24 wherein the muscle tissue is cardiac tissue.

26. The method of claim 24 wherein the muscle tissue is skeletal muscle tissue.

27. A device comprising: one or more processors; and a memory storing instructions which, when executed by the one or more processors, cause the one or more processors to:obtain an image of an artificial tissue under a first set of conditions; extract a first region from the image such that the first region circumscribes the artificial tissue within the image; extract one or more geometric features from the first region, wherein the one or more geometric features provide a shape-based characterization of the artificial tissue; and generate a first tissue descriptor based on the one or more geometric features, wherein the first tissue descriptor encodes a phenotype of the artificial tissue under the first set of conditions.

28. A non-transitory machine readable medium comprising instructions which, when executed by one or more processors, cause the one or more processors to: obtain an image of an artificial tissue under a first set of conditions; extract a first region from the image such that the first region circumscribes the artificial tissue within the image; extract one or more geometric features from the first region, wherein the one or more geometric features provide a shape-based characterization of the artificial tissue; and generate a first tissue descriptor based on the one or more geometric features, wherein the first tissue descriptor encodes a phenotype of the artificial tissue under the first set of conditions.

29. A system for quality control (QC) of artificial tissue, the system comprising: a bioreactor comprising: a device configured for growing tissue; and a sensor assembly configured to obtain one or more images of a tissue within the device; and a QC unit communicatively coupled to the bioreactor, the QC unit comprising one or more processors and a memory storing instructions which, when executed by the one or more processors, cause the one or more processors to: :obtain, from the bioreactor at a first predetermined time point, a first image of the tissue within the device; extract one or more image segments from the first image, wherein each of the one or more image segments comprise a region of interest of the tissue within the first image; determine one or more tissue parameters from the one or more image segments, wherein the one or more tissue parameters are indicative of a physiological state of the tissue at the first predetermined time point; and output a QC report based on the one or more tissue parameters.

30. The system of claim 29 wherein the physiological state comprises one or more of: a maturation state, a presence of disease phenotype, a drug treatment effect, and an irregularity associated with the tissue.

31. The system of claim 30 wherein the physiological state comprises a maturation state and the first predetermined time point is a time point occurring during a maturation period of the tissue.

32. The system of claim 29 wherein the one or more tissue parameters comprise one or more of: a tissue irregularity grading; a geometric parameter; and a passive tension measure.

33. The system of claim 32 wherein the tissue irregularity grading comprises a tissue attachment grading indicative of a degree of attachment of a portion of the artificial tissue to an attachment site of the device.

34. The system of claim 33 wherein the attachment site is one of a tissue scaffold of the device, a side wall of the device near to the tissue scaffold, and a side wall of the device away from the tissue scaffold.

35. The system of claim 33 wherein the tissue attachment grading comprises a plurality of scores including an attachment score, an attachment severity score, and a translucence score.

36. The system of claim 33 wherein the tissue irregularity grading is determined using a grading model based on an image segment of the one or more image segments.

37. The system of claim 36 wherein the image segment comprises a quadrant of the first image which includes the portion of the tissue.

38. The system of claim 36 wherein the grading model comprises a hierarchy of classifiers.

39. The system of claim 38 wherein the grading model comprises a trained multi-task learning (MTL) network.

40. The system of claim 39 wherein the trained MTL network comprises an input convolutional neural network configured to receive an input image, a plurality of independent output networks each configured to estimate a score associated with the input image, and a shared network coupled between the input convolutional neural network and each of the plurality of independent output networks.

41. The system of claim 32 wherein the geometric parameter comprises one of: a tissue length, a tissue width, a tissue volume, or a tissue area.

42. The system of claim 41 wherein the geometric parameter is determined from an image segment of the one or more image segments, the image segment comprising a region of the first image which circumscribes the tissue.

43. The system of claim 42 wherein the tissue length is associated with a length of a major axis of the image segment.

44. The system of claim 43 wherein the tissue width is associated with a predetermined relative position along the major axis of the tissue.

45. The system of claim 44 wherein the tissue width is associated with a length of a minor axis of the image segment at the predetermined relative position.

46. The system of claim 32 wherein the passive tension measure is determined from an image segment of the one or more image segments, the image segment comprising a region of the first image which includes a tissue scaffold of the device.

47. The system of claim 46 wherein the passive tension measure is estimated from a model of displacement of the tissue scaffold.

48. The system of claim 47 wherein the model comprises one or more linear models.

49. The system of claim 48 wherein the model comprises a linear model and a quadratic model.

50. The system of claim 29 wherein the one or more processors of the QC unit are configured to extract the one or more image segments from the first image using a segmentation pipeline.

51. The system of claim 50 wherein the segmentation pipeline includes a plurality of sequential morphological operations.

52. The system of claim 50 wherein the segmentation pipeline includes one or more of: an image equalization operation; an affine transformation; an image crop operation; and a binarization.

53. The system of claim 29 wherein the first image comprises a brightfield microscopy image.

54. The system of claim 29 wherein the QC report includes a tissue quality alert.

55. The system of claim 54 wherein the tissue quality alert is included in the QC report when at least one of the one or more tissue parameters does not meet a predetermined quality control threshold.

56. The system of claim 54 wherein the instructions, when executed by the one or more processors of the QC unit, further cause the one or more processors to: cause a command to be transmitted to the bioreactor based on the tissue quality alert.

57. The system of claim 29 wherein the tissue comprises muscle tissue.

58. The system of claim 57 wherein the muscle tissue is cardiac tissue.

59. The system of claim 57 wherein the muscle tissue is skeletal muscle tissue.

60. A method for quality control (QC) of artificial tissue, the method comprising:obtaining, by one or more processors at a first predetermined time point, a first image of an artificial tissue; extracting, by the one or more processors, one or more image segments from the first image, wherein each of the one or more image segments comprise a region of interest of the artificial tissue within the first image; determining, by the one or more processors, one or more tissue parameters from the one or more image segments, wherein the one or more tissue parameters are indicative of a maturation state of the artificial tissue at the first predetermined time point; and outputting, by the one or more processors, a QC report based on the one or more tissue parameters.

61. A non-transitory machine readable medium comprising instructions which, when executed by one or more processors, cause the one or more processors to: obtain, at a first predetermined time point, a first image of an artificial tissue; extract one or more image segments from the first image, wherein each of the one or more image segments comprise a region of interest of the artificial tissue within the first image; determine one or more tissue parameters from the one or more image segments, wherein the one or more tissue parameters are indicative of a maturation state of the artificial tissue at the first predetermined time point; and output a QC report based on the one or more tissue parameters.

62. A system for engineered tissue attachment grading, the system comprising: a multi-task learning (MTL) network comprising: an input convolutional neural network configured to receive an input image; a plurality of independent output networks each configured to estimate an attachment score associated with the input image; and a shared network coupled between the input convolutional neural network and each of the plurality of independent output networks;a control unit communicatively coupled to the MTL network and comprising one or more processors and computing instructions, which when executed by the one or more processors, cause the one or more processors to: obtain an image of an engineered tissue grown within a device comprising a tissue scaffold for attachment to the engineered tissue; extract a first region from the image, wherein the first region of the image includes a first portion of the engineered tissue and a portion of the tissue scaffold; determine, using the MTL network, a plurality of attachment scores based on the first region; determine an attachment grading for the first region based on the plurality of attachment scores, wherein the attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to an attachment site of the device; and output the attachment grading.

63. The system of claim 62 wherein the attachment site is one of the tissue scaffold, a first portion of a side of a well of the device near to the tissue scaffold, and a second portion of the side of the well of the device away from the tissue scaffold.

64. The system of claim 62 wherein the plurality of attachment scores include an attachment indicator, an attachment severity score, and a translucence score.

65. The system of claim 62 wherein the attachment grading comprises a combination of the plurality of attachment scores.

66. The system of claim 62 wherein the plurality of independent output networks further comprises an auxiliary output network configured to predict the attachment grading.

67. The system of claim 62 wherein the input convolutional neural network comprises a plurality of convolutional neural networks.

68. The system of claim 67 wherein the plurality of convolutional neural networks comprise a plurality of inception blocks.

69. The system of claim 62 wherein each of the plurality of independent output networks comprise a plurality of dense layers and at least one output layer.

70. The system of claim 62 wherein the shared network comprises at least one pooling layer and a plurality of dense layers.

71. The system of claim 62 wherein the first region comprises a first quadrant of the image.

72. The system of claim 62 wherein the computing instructions are further configured, when executed by the one or more processors, to cause the one or more processors to: train the MTL on a training data set comprising a plurality of images of engineered tissues and associated tissue scaffolds, each of the plurality of images associated with a corresponding plurality of attachment scores.

73. The system of claim 62 wherein the image is obtained from a bioreactor within which the engineered tissue is grown.

74. The system of claim 73 further comprising the bioreactor.

75. The system of claim 62 wherein the attachment grading is output as part of an alert.

76. The system of claim 75 wherein the alert is issued when the attachment grading does not meet a predetermined attachment criterion.

77. The system of claim 62 wherein the engineered tissue comprises muscle tissue.

78. The system of claim 77 wherein the muscle tissue is cardiac tissue.

79. The system of claim 77 wherein the muscle tissue is skeletal muscle tissue.

80. A method for engineered tissue attachment grading, the method comprising: obtaining an image of an engineered tissue grown within a device comprising a tissue scaffold for attachment to the engineered tissue;extracting a first region from the image, wherein the first region of the image includes a first portion of the engineered tissue and a portion of the tissue scaffold; determining, using a multi-task learning (MTL) network, a plurality of attachment scores based on the first region; determining an attachment grading for the first region based on the plurality of attachment scores, wherein the attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to an attachment site of the device; and outputting the attachment grading.

81. The method of claim 80 wherein the MTL network comprises an input convolutional neural network configured to receive an input image, a plurality of independent output networks each configured to estimate an attachment score associated with the input image, and a shared network coupled between the input convolutional neural network and each of the plurality of independent output networks.

82. A non-transitory machine readable medium comprising instructions which, when executed by one or more processors, cause the one or more processors to: obtain an image of an engineered tissue grown within a device comprising a tissue scaffold for attachment to the engineered tissue; extract a first region from the image, wherein the first region of the image includes a first portion of the engineered tissue and a portion of the tissue scaffold; determine, using a multi-task learning (MTL) network, a plurality of attachment scores based on the first region; determine an attachment grading for the first region based on the plurality of attachment scores, wherein the attachment grading is indicative of a degree of attachment of the first portion of the engineered tissue to an attachment site of the device; and output the attachment grading.

83. The non-transitory machine readable medium of claim 82 wherein the MTL network comprises an input convolutional neural network configured to receive an input image, a plurality of independent output networks each configured to estimate anattachment score associated with the input image, and a shared network coupled between the input convolutional neural network and each of the plurality of independent output networks.

Images