Organoid brightfield image segmentation method, apparatus, device, and medium

Abstract

Embodiments of the present application provide an organoid bright field image segmentation method, device, equipment and medium, relating to the technical field of image segmentation, including obtaining an original microscopic image of a high-throughput bright field organoid containing a bounding box, and cutting each organoid instance from the original microscopic image into an independent single-instance image according to the bounding box; performing superpixel segmentation on the single-instance image to obtain a plurality of superpixel regions, and extracting skewness parameters and mean parameters based on the intensity distribution of each color channel pixel value in each superpixel region to form skewness Gaussian embedding feature vectors; performing clustering operations on the skewness Gaussian embedding feature vectors as input to obtain boundary class superpixels and non-boundary class superpixels to generate an initial boundary binary image; generating an initial mask proposal according to the initial boundary binary image and a boundary completion algorithm; and training according to the initial mask proposal and a segmentation network to output a final pixel-level segmentation mask. The embodiments of the present application can improve the accuracy and robustness of image segmentation.

Description

Technical Field

[0001] This application relates to the field of image segmentation technology, and in particular to a method, apparatus, device and medium for bright-field image segmentation of organoids. Background Technology

[0002] In related technologies, patient-derived organoids (PDOs) can reproduce tumor tissue features in vitro. High-throughput culture generates a large number of bright-field images, requiring pixel-level segmentation of each organoid to support morphological analysis and drug efficacy evaluation. Deep learning semantic segmentation (such as U-Net) is highly effective, but it relies on massive amounts of fine-grained mask annotations. However, each PDO image contains hundreds of instances, requiring specialized expertise for annotation and extremely time-consuming boundary delineation, severely limiting its large-scale application. To reduce annotation costs, existing solutions generate pseudo-labels using bounding boxes: BoxSup generates approximately 2000 candidate masks through multi-scale combination grouping and iterative training, but the computational cost is enormous; GrabCut extracts the foreground based on graph cuts, is sensitive to boundary contrast, and has limited accuracy under the complex textures of organoids; traditional image processing methods do not require training, but are not robust to noise and boundary discontinuities, making them difficult to adapt to real-world scenarios with diverse organoid morphologies and complex boundaries. Summary of the Invention

[0003] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a method, apparatus, device, and medium for bright-field image segmentation of organoids, aiming to improve the accuracy and robustness of image segmentation.

[0004] In a first aspect, embodiments of this application provide a method for bright-field image segmentation of organoids, the method comprising: Obtain the original microscopic image of a high-throughput bright-field organoid containing a bounding box, and crop each organoid instance from the original microscopic image into an independent single-instance image based on the bounding box; The single instance image is segmented into multiple superpixel regions, and the skewness parameter and mean parameter are extracted based on the intensity distribution of each color channel pixel value in each superpixel region to form a skewed Gaussian embedding feature vector. Clustering is performed using the skewed Gaussian embedding feature vector as input to obtain boundary class superpixels and non-boundary class superpixels to generate an initial boundary binary image; Based on the initial boundary binary image and boundary completion algorithm, an initial mask proposal is generated; The network is trained based on the initial mask proposal and the segmentation network to output the final pixel-level segmentation mask.

[0005] According to some embodiments of this application, the step of performing superpixel segmentation on the single instance image to obtain multiple superpixel regions, and extracting skewness parameters and mean parameters based on the intensity distribution of pixel values ​​of each color channel within each superpixel region to construct a skewed Gaussian embedding feature vector, includes: The single instance image is segmented into multiple superpixel regions with similar texture and color features by using a simple linear iterative clustering algorithm. For each superpixel region, extract the intensity distribution of pixel values ​​of each color channel within the superpixel region and fit a skewed Gaussian distribution; The skewness parameter and mean parameter are obtained from the skewed Gaussian distribution to form the skewed Gaussian embedding feature vector.

[0006] According to some embodiments of this application, the step of performing superpixel segmentation on the single instance image based on a simple linear iterative clustering algorithm to obtain multiple superpixel regions with similar texture and color features includes: A five-dimensional space is established, consisting of a color space and two-dimensional coordinates, wherein the color space includes the RGB color space or the CIELAB color space; Obtain the superpixel count parameter at multiple scales; In the five-dimensional space, the single instance image is segmented into multiple superpixel regions with similar texture and color features by using a simple linear iterative clustering algorithm.

[0007] According to some embodiments of this application, generating an initial mask proposal based on the initial boundary binary image and the boundary completion algorithm includes: The centroid coordinates of the organoid are obtained from the initial boundary binary image; According to the boundary completion algorithm, rays are emitted from the centroid coordinates of the organoid to the sampling points of the four sides of the bounding box to detect the boundary pixels in each direction. The direction where there are no boundary pixels or the boundary pixels change abnormally is completed to generate a closed boundary image. Perform region filling on the closed boundary image to generate an initial mask proposal.

[0008] According to some embodiments of this application, the step of emitting rays from the centroid coordinates of the organoid to the sampling points of the four sides of the bounding box according to the boundary completion algorithm to detect boundary pixels in each direction, and completing the boundary pixels in directions where there are no boundary pixels or abnormal abrupt changes in boundary pixels to generate a closed boundary image includes: According to the boundary completion algorithm, rays are emitted from the centroid coordinates of the organoid to the sampling points on the four sides of the bounding box to detect boundary pixels in each direction. When the boundary pixel exists, the boundary radius of the boundary pixel is recorded as the effective boundary radius; When the boundary pixel is missing or undergoes an abnormal change, it is filled in according to the radius statistics of adjacent effective directions to obtain the effective boundary radius; Based on the effective boundary radius, generate multiple candidate masks; Multiple candidate masks are fused by voting to generate a closed boundary image.

[0009] According to some embodiments of this application, the step of performing region filling on the closed boundary image to generate multiple candidate masks includes: Obtain the centroid coordinates of the target within the region enclosed by the closed boundary image; Using the target centroid coordinates as the seed point, a flood fill algorithm is executed to set all pixels within the closed contour of the closed boundary image as foreground values, generating multiple candidate masks. The flood fill algorithm starts from the seed point and spreads inward in a predetermined number of neighborhoods until it encounters the boundary pixels of the closed boundary image and stops spreading.

[0010] According to some embodiments of this application, the step of training based on the initial mask proposal and the segmentation network to output a final pixel-level segmentation mask includes: Using the initial mask proposal as a pseudo-label, the segmentation network is iteratively trained to obtain the trained segmentation network. After each training round, the predicted mask generated by the trained segmentation network is obtained, and pseudo-labels are selected or fused from the initial mask proposal and the predicted mask according to the dynamic decay ratio to generate pseudo-labels for the next training round. The process continues until the preset number of iterations is reached, at which point the final pixel-level segmentation mask is output.

[0011] Secondly, embodiments of this application provide an organoid bright-field image segmentation apparatus, the apparatus comprising: The acquisition module is used to acquire the original microscopic image of a high-throughput bright-field organoid containing a bounding box, and to crop each organoid instance from the original microscopic image into an independent single-instance image based on the bounding box. The segmentation module is used to perform superpixel segmentation on the single instance image to obtain multiple superpixel regions, and extract skewness parameters and mean parameters based on the intensity distribution of pixel values ​​of each color channel in each superpixel region to form a skewed Gaussian embedding feature vector. The clustering operation module is used to perform clustering operations with the skewed Gaussian embedding feature vector as input to obtain boundary class superpixels and non-boundary class superpixels to generate an initial boundary binary image. The generation module is used to generate an initial mask proposal based on the initial boundary binary image and the boundary completion algorithm; The output module is used to train the network based on the initial mask proposal and the segmentation network, and output the final pixel-level segmentation mask.

[0012] Thirdly, embodiments of this application provide a computer device, including: Memory, used to store programs; A processor for executing a program stored in the memory, wherein when the processor executes the program stored in the memory, the processor is configured to perform the method described in the first aspect above. Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions for performing the method described in the first aspect above.

[0013] According to the technical solution of the embodiments of this application, at least the following beneficial effects are achieved: The embodiments of this application acquire original microscopic images of high-throughput bright-field organoids containing bounding boxes; each organoid instance is cropped from the original microscopic image into an independent single-instance image based on the bounding boxes; superpixel segmentation is performed on the single-instance images to obtain multiple superpixel regions; skewness parameters and mean parameters are extracted based on the intensity distribution of pixel values ​​in each color channel within each superpixel region to construct a skewed Gaussian embedding feature vector; clustering is performed using the skewed Gaussian embedding feature vector as input to obtain boundary superpixels and non-boundary superpixels to generate an initial boundary binary image; an initial mask proposal is generated based on the initial boundary binary image and a boundary completion algorithm; the initial mask proposal and a segmentation network are trained to output a final pixel-level segmentation mask. This application's embodiments extract skewness and mean parameters of pixel intensity distribution or color channel intensity distribution within superpixel regions to construct skewed Gaussian embedding features. These features characterize the asymmetric intensity distribution differences typically present in bright-field images of organoid boundaries, interiors, and background regions, enabling boundary-class superpixels and non-boundary-class superpixels to be separable in feature space. This overcomes the problem of inaccurate boundary localization in traditional color similarity models. Simultaneously, boundary responses are detected based on rays emitted from the centroid to sampling points in multiple directions, and missing or discontinuous boundary directions are filled in. This allows for the recovery of relatively complete instance contours without additional pixel-level annotations, generating high-quality initial mask proposals. Furthermore, a dynamically decaying hybrid pseudo-label refinement mechanism is used to gradually adjust the ratio of initial mask proposals to network-predicted masks during iterative training, reducing the risk of erroneous pseudo-label accumulation and enhancing cross-dataset segmentation stability, thereby improving the accuracy and robustness of image segmentation.

[0014] The solutions provided in the second to fourth aspects above are used to implement or cooperate with the methods provided in the first aspect above, and therefore can achieve the same or corresponding beneficial effects as the first aspect, which will not be elaborated here.

[0015] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0016] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0017] Figure 1 A flowchart illustrating an organoid bright-field image segmentation method provided in one embodiment of this application; Figure 2 A schematic diagram illustrating the process of constructing a skewed Gaussian embedding feature vector according to one embodiment of this application; Figure 3 A schematic diagram illustrating the process of obtaining multiple superpixel regions with similar texture and color features, provided for one embodiment of this application; Figure 4 A schematic diagram illustrating the process of generating an initial mask proposal provided in one embodiment of this application; Figure 5 A schematic diagram illustrating the process of generating a closed boundary image according to one embodiment of this application; Figure 6 This is a schematic diagram illustrating the process of generating multiple candidate masks according to one embodiment of this application; Figure 7 A schematic diagram illustrating the output of the final pixel-level segmentation mask provided in one embodiment of this application; Figure 8 This is a schematic diagram of the structure of a U-Net network provided in one embodiment of this application; Figure 9 This is a schematic diagram of an organoid bright-field image segmentation apparatus provided in one embodiment of this application; Figure 10 This is a schematic diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical methods, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. It should be noted that the meaning of "multiple" (or "more than") in the description of the embodiments of this application refers to two or more, and "greater than," "less than," "exceeding," etc. are understood to exclude the number itself, while "above," "below," "within," etc. are understood to include the number itself. If "first," "second," etc. are used in the description, they are only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated. In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: the existence of a alone, the existence of b alone, the existence of c alone, the simultaneous existence of a and b, the simultaneous existence of a and c, the simultaneous existence of b and c, or the simultaneous existence of a, b, and c, where a, b, and c can be single or multiple.

[0019] In the description of this application, unless otherwise expressly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0020] In some cases, patient-derived organoids (PDOs) can reproduce tumor tissue features in vitro. High-throughput culture generates a large number of bright-field images, requiring pixel-level segmentation of each organoid to support morphological analysis and efficacy evaluation. Deep learning semantic segmentation (such as U-Net) is highly effective, but it relies on massive amounts of fine-grained mask annotations. However, each PDO image contains hundreds of instances, requiring specialized expertise for annotation and extremely time-consuming boundary delineation, severely limiting its large-scale application. To reduce annotation costs, existing solutions generate pseudo-labels using bounding boxes: BoxSup generates approximately 2000 candidate masks through multi-scale combination grouping and iterative training, but the computational cost is enormous; GrabCut extracts the foreground based on graph cuts, is sensitive to boundary contrast, and has limited accuracy under complex organoid textures; traditional image processing methods do not require training, but are not robust to noise and boundary discontinuities, making them difficult to adapt to real-world scenarios with diverse organoid morphologies and complex boundaries.

[0021] Based on the above, this application proposes a method, apparatus, device, and medium for bright-field image segmentation of organoids, aiming to improve the accuracy and robustness of image segmentation.

[0022] To facilitate understanding of the solutions in the embodiments of this application, the relevant concepts involved in the embodiments of this application will be introduced below.

[0023] SLIC (Simple Linear Iterative Clustering) is an efficient image preprocessing algorithm for generating superpixels. It converts an image from RGB color space to CIELAB color space and constructs a five-dimensional feature vector using pixel spatial coordinates (x, y). It then performs K-means iterative clustering within a finite local region. By introducing a compactness parameter to balance color similarity and spatial proximity, it generates superpixel blocks that are regular in shape, compact, and closely attached to the natural boundaries of objects. In medical image segmentation, SLIC is often used as a pre-processing step for deep learning models such as U-Net, compressing tens of thousands of original pixels into a controllable number of superpixels, significantly reducing the computational burden of subsequent feature processing while preserving crucial edge structure information, thus improving the model's ability to perceive local textures.

[0024] BCA (Boundary Completion Algorithm) is a key restoration technique in image segmentation post-processing, designed to eliminate breaks, burrs, and internal holes at the edges of a target image, generating a continuous, smooth, and complete closed contour. This algorithm effectively addresses edge discontinuities caused by uneven illumination, blurred tissue texture, or prediction errors in depth models. Introducing boundary completion after the output of segmentation models such as U-Net can significantly improve the smoothness and integrity of the mask, which is crucial for downstream tasks such as accurate clinical measurements and 3D surface reconstruction.

[0025] K-means clustering is a classic unsupervised hard clustering algorithm commonly used in traditional image segmentation for initial pixel-level classification. The algorithm first randomly initializes K cluster centers, then iteratively executes two steps: an assignment step, which calculates the feature distance (usually grayscale or RGB Euclidean distance) from each pixel to each center and assigns it to the nearest neighbor cluster; and an update step, which recalculates the mean of all pixels within each cluster as the new center, until the center points converge. In segmentation scenarios, it divides the image into K non-overlapping regions based on the similarity of pixel features.

[0026] Flood Fill (Flood Fill Algorithm) is a classic connected component filling algorithm based on region growing, widely used in interactive image editing and post-segmentation processing. Starting from a specified seed pixel, it recursively or iteratively searches and labels all neighboring pixels within its four- or eight-neighborhood that meet similarity criteria, based on a preset color or grayscale tolerance threshold, until expansion is impossible. In image segmentation, this algorithm has significant post-processing value: for binary masks output by models such as U-Net, it can quickly locate and fill tiny holes inside the target (hole filling), while removing isolated noise points outside the main region, thus obtaining semantically coherent and complete connected regions.

[0027] IoU (Intersection over Union) is the most core and commonly used metric for evaluating image segmentation models. It calculates the degree of overlap between the predicted region and the ground truth region.

[0028] The organoid bright-field image segmentation method provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms; the software can be an application that implements the organoid bright-field image segmentation method, but is not limited to the above forms. This application can be applied to numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via communication networks. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices. It should be noted that in various specific embodiments of this invention, when processing is required based on data related to the characteristics of an object (e.g., user attributes or sets of attribute information), permission or consent from the corresponding object is obtained first, and the collection, use, and processing of this data comply with relevant laws and standards. Furthermore, when the embodiments of the present invention need to obtain the attribute information of an object, they will obtain the separate permission or separate consent of the corresponding object through pop-up windows or redirection to a confirmation page. After obtaining the separate permission or separate consent of the corresponding object, they will then obtain the relevant data of the object necessary for the embodiments of the present invention to operate normally.

[0029] See Figure 1 , Figure 1 This is a flowchart illustrating an organoid bright-field image segmentation method according to an embodiment of this application. The organoid bright-field image segmentation method provided in this embodiment includes, but is not limited to, steps S110 to S150, which are described below. Step S110: Obtain the original microscopic image of the high-throughput bright-field organoid containing the bounding box, and crop each organoid instance from the original microscopic image into an independent single-instance image according to the bounding box; Step S120: Perform superpixel segmentation on the single instance image to obtain multiple superpixel regions, and extract skewness parameters and mean parameters based on the intensity distribution of pixel values ​​of each color channel in each superpixel region to form a skewed Gaussian embedding feature vector. Step S130: Perform clustering operation with skewed Gaussian embedding feature vector as input to obtain boundary class superpixels and non-boundary class superpixels to generate an initial boundary binary image; Step S140: Generate an initial mask proposal based on the initial boundary binary image and the boundary completion algorithm; Step S150: Train the network based on the initial mask proposal and the segmentation network to output the final pixel-level segmentation mask.

[0030] It should be noted that the original microscopic images contain 100 to 600 organoid instances, and directly processing the entire image would be computationally too expensive. First, based on the manually annotated bounding box coordinates, each organoid instance is cropped from the original image into an independent single-instance image, which serves as the unified input for subsequent modules. The cropped images are then uniformly scaled to 128×128 pixels and converted to RGB three-channel format. The experimental dataset in this embodiment of the application yielded 3283 single-instance images cropped from 18 full-field bright-field images.

[0031] It should be noted that the skewness statistical features were quantitatively validated by independently dividing the boundary, interior, and background superpixels using the real mask. Table 1 shows the skewness statistics for superpixel categories. Among them, boundary superpixels and interior superpixels showed stable differences in mean and median skewness, and the Mann-Whitney U test results showed that the difference was extremely significant; similarly, there was a significant difference between interior and background superpixels. The validation process was based on independently defining categories according to the spatial overlap relationship of the real mask, avoiding circular reasoning. This proves that the skewness parameter is not an arbitrarily chosen empirical feature, but rather a statistical feature corresponding to the bright-field organoid boundary imaging mechanism and capable of producing actual discriminative effects on boundary detection. This allows for the acquisition of relatively reliable boundary candidate regions under weak supervision, providing high-quality initial information for subsequent boundary completion and pseudo-label refinement.

[0032] Table 1 Superpixel Category Skewness Statistics

[0033] It should be noted that, Using the skewed Gaussian embedding feature vectors of each superpixel as input, K-means clustering is performed to divide the superpixels into two groups: boundary classes and non-boundary classes. After clustering, the boundary class is determined based on the skewness distribution, mean value, and consistency with the physical laws of bright-field imaging of the cluster centers, rather than relying solely on a fixed grayscale threshold. The resulting initial boundary binary image can adapt to PDO bright-field images of different batches, brightness levels, and shapes, reducing the parameter sensitivity of thresholding and ordinary edge detection methods across image scenes.

[0034] In one embodiment, a high-throughput bright-field organoid image containing bounding boxes is acquired. Each organoid instance is cropped from the original microscopic image into an independent single-instance image based on the bounding boxes. Superpixel segmentation is performed on the single-instance image to obtain multiple superpixel regions. The skewness and mean parameters of the pixel grayscale intensity distribution within each superpixel region are extracted to form a skewed Gaussian embedding feature vector. A clustering operation is performed using the skewed Gaussian embedding feature vector as input to obtain boundary superpixels and non-boundary superpixels to generate an initial boundary binary image. An initial mask proposal is generated based on the initial boundary binary image and a boundary completion algorithm. The initial mask proposal and a segmentation network are trained to output the final pixel-level segmentation mask. This application's embodiments extract skewness and mean parameters of pixel intensity distribution or color channel intensity distribution within superpixel regions to construct skewed Gaussian embedding features. These features characterize the asymmetric intensity distribution differences typically present in bright-field images of organoid boundaries, interiors, and background regions, enabling boundary-class superpixels and non-boundary-class superpixels to be separable in feature space. This overcomes the problem of inaccurate boundary localization in traditional color similarity models. Simultaneously, boundary responses are detected based on rays emitted from the centroid to sampling points in multiple directions, and missing or discontinuous boundary directions are filled in. This allows for the recovery of relatively complete instance contours without additional pixel-level annotations, generating high-quality initial mask proposals. Furthermore, a dynamically decaying hybrid pseudo-label refinement mechanism is used to gradually adjust the ratio of initial mask proposals to network-predicted masks during iterative training, reducing the risk of erroneous pseudo-label accumulation and enhancing cross-dataset segmentation stability, thereby improving the accuracy and robustness of image segmentation.

[0035] See Figure 2 , Figure 2 This is a schematic diagram illustrating the process of constructing a skewed Gaussian embedding feature vector according to an embodiment of this application. Regarding step S120, which involves superpixel segmentation of a single instance image to obtain multiple superpixel regions, and extracting skewness and mean parameters based on the intensity distribution of pixel values ​​in each color channel within each superpixel region to construct a skewed Gaussian embedding feature vector, this includes, but is not limited to, steps S210 to S230. Each step will be described in detail below. Step S210: Perform superpixel segmentation on a single instance image based on a simple linear iterative clustering algorithm to obtain multiple superpixel regions with similar texture and color features; Step S220: For each superpixel region, extract the intensity distribution of each color channel pixel value within the superpixel region and fit a skewed Gaussian distribution; Step S230: Obtain the skewness parameter and mean parameter based on the skewed Gaussian distribution to form a skewed Gaussian embedding feature vector.

[0036] In one embodiment, the cropped single-instance image is segmented into superpixel regions using the Simple Linear Iterative Clustering (SLIC) algorithm, dividing the image into several superpixel regions with similar texture and color features. For each superpixel region, the intensity distribution of pixel values ​​in each color channel within the region is extracted and fitted with a skewed Gaussian distribution. The probability density function of the skewed Gaussian distribution is:

[0037] in, It is the standard normal probability density function. The standard normal cumulative distribution function is... For location parameters (mean). For scale parameters, This is a skewness parameter, reflecting the degree and direction of asymmetry in the distribution.

[0038] It should be noted that a two-dimensional feature vector for each superpixel is constructed using the skewness parameter and the mean parameter, serving as the skewed Gaussian embedding representation for that region. The mean parameter characterizes the overall brightness level of the superpixel, while the skewness parameter characterizes the asymmetry direction and long tail of the intensity distribution within the superpixel. By performing skewed modeling at the superpixel level, interference from pixel-by-pixel noise on boundary judgment can be avoided, and the ability to identify semi-transparent, weak-contrast, and discontinuous boundaries can be improved. Further ablation experiments show that, while maintaining consistency with subsequent processes such as BCA, if only Otsu thresholding is used for superpixel classification without skewed Gaussian embedding, the IoU at the candidate mask stage is 0.412; using standard Gaussian distribution fitting, the IoU is 0.383; while using the skewed Gaussian embedding of this invention, the IoU is improved to 0.677, representing improvements of 26.5% and 29.4%, respectively.

[0039] See Figure 3 , Figure 3 This is a schematic diagram illustrating the process of obtaining multiple superpixel regions with similar texture and color features according to one embodiment of this application. Regarding step S210, which uses a simple linear iterative clustering algorithm to perform superpixel segmentation on a single instance image to obtain multiple superpixel regions with similar texture and color features, this includes, but is not limited to, steps S310 to S330. Each step will be described in detail below.

[0040] Step S310: Establish a five-dimensional space consisting of a color space and two-dimensional coordinates, wherein the color space includes RGB color space or CIELAB color space. Step S320: Obtain the superpixel count parameters at multiple scales; Step S330: Perform superpixel segmentation on a single instance image in five-dimensional space based on a simple linear iterative clustering algorithm to obtain multiple superpixel regions with similar texture and color features.

[0041] In one embodiment, the SLIC algorithm performs K-means clustering in a five-dimensional space (RGB color space plus two-dimensional spatial coordinates), pixel With cluster center The distance between them is defined as:

[0042] in, The color Euclidean distance between the pixel and the cluster center in RGB space. The spatial Euclidean distance between the pixel and the cluster center. A scale factor to balance the effects of color distance and spatial distance. Superpixel count parameter. The value is taken at multiple scales, with a default base value of 50. The process is repeated at 8 different scaling ratios (0.2, 0.4, 0.5, 1, 1.5, 2, 2.5, 3) to obtain multiple superpixel regions with similar texture and color features.

[0043] See Figure 4 , Figure 4 This is a schematic diagram of the process for generating an initial mask proposal according to an embodiment of this application. The above step S140, which generates the initial mask proposal based on the initial boundary binary image and the boundary completion algorithm, includes, but is not limited to, steps S410 to S440. Each step will be described in turn below.

[0044] Step S410: Obtain the centroid coordinates of the organoid based on the initial boundary binary image; Step S420: Based on the boundary completion algorithm, rays are emitted from the centroid coordinates of the organoid to the sampling points on the four sides of the bounding box to detect the boundary pixels in each direction. For directions where there are no boundary pixels or abnormal abrupt changes in boundary pixels, the boundary is completed to generate a closed boundary image. Step S430: Perform region filling on the closed boundary image to generate multiple candidate masks; Step S440: Perform voting and fusion on multiple candidate masks to generate an initial mask proposal.

[0045] In one embodiment, using the binary image of the boundary obtained by skewed Gaussian embedding clustering as input, the centroid of the target region is first calculated, and then rays are uniformly emitted from the centroid to the sampling points on the four sides of the bounding box. The existence of boundary pixels is detected along each ray, and the effective boundary radius is recorded. For directions where there is no boundary response or where the radius undergoes an abrupt change, these are identified as boundary gap directions, and the gaps are filled using the radius statistics of adjacent effective directions or the global average radius. Finally, a closed boundary is obtained and flooding filling is performed, outputting multiple candidate masks. These candidate masks are then fused through voting to generate an initial mask proposal. This transforms the initial mask from a scattered boundary response into a fillable closed contour. The completion direction of the boundary completion algorithm is constrained by both the instance centroid and the bounding box sampling points, enabling directional sealing for directions without boundary response and reducing shape expansion caused by blind expansion.

[0046] See Figure 5 , Figure 5 This is a schematic flowchart illustrating the generation of a closed boundary image according to an embodiment of this application. Regarding step S420, which involves emitting rays from the centroid coordinates of the organoid to the sampling points on the four sides of the bounding box using a boundary completion algorithm to detect boundary pixels in each direction, and completing the boundary images in directions where no boundary pixels exist or where boundary pixels abruptly change, the process includes, but is not limited to, steps S510 to S540. Each step will be described in detail below.

[0047] Step S510: Based on the boundary completion algorithm, rays are emitted from the centroid coordinates of the organoid to the sampling points on the four sides of the bounding box to detect boundary pixels in each direction. Step S520: When boundary pixels exist, record the boundary radius of the boundary pixels as the effective boundary radius; Step S530: When the boundary pixel is missing or an abnormal change occurs, the effective boundary radius is obtained by filling in the missing pixels based on the radius statistics of adjacent effective directions. Step S540: Generate a closed boundary image based on the effective boundary radius.

[0048] In one embodiment, the boundary completion algorithm includes centroid localization, bounding box sampling point generation, ray scanning and gap detection, and average radius estimation and gap filling.

[0049] Step 1: Centroid localization. For the binary image of the boundary... The centroid coordinates of the organoid are obtained by calculating the moments of the maximum contour. .

[0050] Step 2: Generate bounding box sampling points. Along the bounding box... The set of uniformly sampled points along the four sides , as the endpoint of the ray.

[0051] Step 3: X-ray scanning and notch detection. For each sampling point... Construction from the center of mass arrive line segments Sample all pixel coordinates along the line segment and check for the existence of boundary pixels (i.e., (Pixels with a value of 1): If there are no boundary pixels on the line segment, it is determined that there is a boundary gap in that direction, and... Add to candidate completion line segment set If boundary pixels exist, record the maximum distance from the boundary pixel to the centroid. Add to distance set .

[0052] Step 4: Mean Radius Estimation and Gap Filling. Calculate the mean boundary radius. .right Each candidate line segment The distance from the centroid of the line segment does not exceed The pixel is set to the boundary pixel ( Complete the gap filling and output the closed boundary image. .

[0053] See Figure 6 , Figure 6 This is a schematic diagram illustrating the process of generating multiple candidate masks according to one embodiment of this application. Regarding step S430, which involves filling the closed boundary image to generate multiple candidate masks, steps S610 to S620 are included, but are not limited to. Each step will be described in detail below.

[0054] Step S610: Obtain the centroid coordinates of the target within the region enclosed by the closed boundary image; Step S620: Using the target centroid coordinates as the seed point, execute the flood fill algorithm to set all pixels inside the closed contour of the closed boundary image as foreground values ​​and generate multiple candidate masks. The flood fill algorithm starts from the seed point and spreads inward in a preset number of neighborhoods until it encounters the boundary pixels of the closed boundary image and stops spreading.

[0055] In one embodiment, the completed closed boundary image is processed using centroid coordinates. Using the seed point, the Flood Fill algorithm is called to fill inwards, setting all pixels within the closed contour as foreground (with a value of 1), and generating an initial binary mask proposal. When there are gaps or gaps at the boundary, relying solely on conventional Flood Fill may lead to region leakage. Boundary completion algorithms, however, can first close the gaps before performing region filling, thereby improving the integrity and stability of the candidate mask. After excluding small-sized blurred images, the boundary completion algorithm achieves the highest average IoU as a candidate mask generation strategy, significantly outperforming other filling strategies.

[0056] See Figure 7 , Figure 7 This is a schematic diagram illustrating the process of outputting the final pixel-level segmentation mask according to an embodiment of this application. The above step S150, which trains the segmentation network based on the initial mask proposal and outputs the final pixel-level segmentation mask, includes, but is not limited to, steps S710 to S730, which will be described in detail below.

[0057] Step S710: Using the initial mask proposal as pseudo-labels, iteratively train the segmentation network to obtain the trained segmentation network; Step S720: After each round of training, obtain the predicted mask generated by the segmentation network after training, and select or fuse pseudo-labels from the initial mask proposal and the predicted mask according to the dynamic decay ratio to generate pseudo-labels for the next round of training. Step S730: Continue until the number of iterations for training reaches the preset number of iterations, then output the final pixel-level segmentation mask.

[0058] It should be noted that the high-quality single candidate mask generated by the aforementioned skewed Gaussian embedding and boundary completion algorithms is used as a source of weakly supervised pseudo-labels, and a dynamically decaying hybrid pseudo-label mechanism drives U-Net to correct boundary errors round by round. If GrabCut candidate masks are used for the same iterative U-Net training, the final IoU is only 0.060, lower than the IoU of GrabCut candidate masks themselves (0.225), indicating that low-quality pseudo-labels can mislead the network to learn incorrect patterns. Therefore, this application first generates high-quality single candidate masks suitable for PDO bright-field images, and then refines them using U-Net.

[0059] See Figure 8 , Figure 8 This is a schematic diagram of the structure of a U-Net network provided in one embodiment of this application.

[0060] It should be noted that, from left to right, the components are the encoder (downsampling path), bottleneck layer, decoder (upsampling path), and skip connections. The names of each component are labeled as follows: Encoder: Consists of 5 downsampling blocks ( to It consists of two consecutive downsampling blocks. Convolutional layers and the Rectified Linear Activation Function (ReLU). to Before convolution, Downsampling is performed using the max pooling layer. The number of feature map channels in each stage are 64, 128, 256, 512, and 1024, respectively.

[0061] Bottleneck layer: The output feature map size is (standard (Enter below).

[0062] The decoder consists of four upsampling stages. The first stage uses a single transposed convolution ( (Step size 2) reduces the number of channels from 1024 to 512; the subsequent three upsampling stages ( to Separately concatenate the corresponding encoder skip connection features, and then pass through two... After the convolutional layer, a transposed convolutional upsampling is performed.

[0063] Output layer: after final The convolution outputs dual-channel logical values ​​(background and foreground), corresponding to binary pixel prediction.

[0064] Network input: RGB single-instance cropped image.

[0065] Model size: 34.51M trainable parameters, 16.34 GMACs (32.68 GFLOPs) computation per forward inference.

[0066] In one embodiment, the generated initial mask proposal As pseudo-labels, U-Net was trained for the first round, a total of 50 epochs, using binary cross-entropy loss as the loss function.

[0067] in Total number of pixels For the first The true label (0 or 1) of each pixel. The model predicts the probability that the pixel belongs to the foreground.

[0068] In one embodiment, Hybrid label updates involve U-Net generating predicted masks for the training set images after each training round, and then weighting and fusing them with the initial mask proposals according to a mixing ratio to generate the labels for the next round of training. In the early stages of training, candidate masks based on physical imaging features and geometric completion are primarily used to avoid premature self-reinforcement errors when the model's pseudo-label quality is insufficient. In the later stages of training, the weights of the model's prediction results are gradually increased, enabling the network to learn morphological patterns across samples and correct local errors in the initial proposals. This establishes a smooth transition between physically prior candidate masks and data-driven predictions to reduce the risk of accumulating erroneous pseudo-labels over multiple training rounds. The hybrid update format is as follows:

[0069] Mixing ratio From initial value Initially, at a fixed attenuation rate Decreasing in successive rounds:

[0070] It should be noted that, with each iteration, the weight of U-Net's own predictions gradually increases, while the weight of the initial proposals gradually decreases. This reduces the risk of accumulating false labels during the iteration process and improves the consistency of the final output mask. Sensitivity analysis shows that, over a wide range of initial mixing ratios, the final IoU fluctuates only within approximately 1.0%.

[0071] In one embodiment, iterative convergence involves 20 rounds of training. Compared to one-shot training, iterative training maintains a high level of stability in terms of mean metrics and further corrects local holes, boundary burrs, and small-scale structural defects in the initial candidate mask. Therefore, the iterative refining module of this application solves the problems of stability, boundary smoothness, and batch consistency in weakly supervised pseudo-label training.

[0072] It should be noted that skewed Gaussian embedding provides boundary candidates based on the physical imaging laws of bright field. BCA completes the candidate boundaries into closed contours based on the geometric relationship between the bounding box and the centroid. Flood Fill transforms the closed contours into initial instance masks, and iterative U-Net then transforms these initial masks into stable, generalizable segmentation models. This combination enables bounding box annotations to be effectively converted into pixel-level instance supervision signals. The final performance of the system on the colorectal PDO dataset (3283 instances from 18 full-view images) is as follows: Intersection over Union (IoU) = 0.804, Dice coefficient = 0.888, precision = 0.865, recall = 0.923, and boundary F-score = 0.205, representing a 57.9% improvement in IoU compared to the GrabCut baseline and a 21.5% improvement in IoU compared to the active contour model. Ablation experiments at the candidate mask stage further demonstrate that skewed Gaussian embedding brings IoU improvements of 26.5% and 29.4% compared to Otsu thresholding classification and standard Gaussian fitting, respectively.

[0073] like Figure 9 As shown, Figure 9 This is a schematic diagram of an organoid bright-field image segmentation apparatus according to an embodiment of this application. The organoid bright-field image segmentation apparatus 900 provided in this embodiment includes: The acquisition module 910 is used to acquire the original microscopic image of the high-throughput bright-field organoid containing the bounding box, and to crop each organoid instance from the original microscopic image into an independent single-instance image based on the bounding box. The segmentation module 920 is used to perform superpixel segmentation on a single instance image to obtain multiple superpixel regions, and extract skewness parameters and mean parameters based on the intensity distribution of pixel values ​​of each color channel in each superpixel region to form a skewed Gaussian embedding feature vector. Clustering operation module 930 is used to perform clustering operations with skewed Gaussian embedding feature vectors as input to obtain boundary class superpixels and non-boundary class superpixels to generate an initial boundary binary image; The generation module 940 is used to generate an initial mask proposal based on the initial boundary binary image and the boundary completion algorithm; Output module 950 is used to train the segmentation network based on the initial mask proposal and output the final pixel-level segmentation mask.

[0074] In one embodiment, a high-throughput bright-field organoid image containing bounding boxes is acquired. Each organoid instance is cropped from the original microscopic image into an independent single-instance image based on the bounding boxes. Superpixel segmentation is performed on the single-instance image to obtain multiple superpixel regions. Skewness and mean parameters are extracted based on the intensity distribution of pixel values ​​in each color channel within each superpixel region to form a skewed Gaussian embedding feature vector. Clustering is performed using the skewed Gaussian embedding feature vector as input to obtain boundary superpixels and non-boundary superpixels to generate an initial boundary binary image. An initial mask proposal is generated based on the initial boundary binary image and a boundary completion algorithm. The initial mask proposal and a segmentation network are trained to output the final pixel-level segmentation mask. This application's embodiments extract skewness and mean parameters of pixel intensity distribution or color channel intensity distribution within superpixel regions to construct skewed Gaussian embedding features. These features characterize the asymmetric intensity distribution differences typically present in bright-field images of organoid boundaries, interiors, and background regions, enabling boundary-class superpixels and non-boundary-class superpixels to be separable in feature space. This overcomes the problem of inaccurate boundary localization in traditional color similarity models. Simultaneously, boundary responses are detected based on rays emitted from the centroid to sampling points in multiple directions, and missing or discontinuous boundary directions are filled in. This allows for the recovery of relatively complete instance contours without additional pixel-level annotations, generating high-quality initial mask proposals. Furthermore, a dynamically decaying hybrid pseudo-label refinement mechanism is used to gradually adjust the ratio of initial mask proposals to network-predicted masks during iterative training, reducing the risk of erroneous pseudo-label accumulation and enhancing cross-dataset segmentation stability, thereby improving the accuracy and robustness of image segmentation.

[0075] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0076] Figure 10 This application illustrates a computer device 1000 provided in an embodiment of the present application. The computer device 1000 may be a server or a terminal, and its internal structure includes, but is not limited to: Memory 1010 is used to store programs; The processor 1020 is used to execute the program stored in the memory 1010. When the processor 1020 executes the program stored in the memory 1010, the processor 1020 is used to execute the above-mentioned organoid bright field image segmentation method. The processor 1020 and the memory 1010 can be connected via a bus or other means. The memory 1010, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs, such as the organoid bright-field image segmentation method described in any embodiment of this application. The processor 1020 implements the above-described organoid bright-field image segmentation method by running the non-transitory software program and instructions stored in the memory 1010. The memory 1010 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function. The data storage area may store the organoid bright-field image segmentation method described above. Furthermore, the memory 1010 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 1010 may optionally include memory remotely located relative to the processor 1020, and these remote memories can be connected to the processor 1020 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The non-transient software program and instructions required to implement the above-described organoid bright-field image segmentation method are stored in memory 1010. When executed by one or more processors 1020, the organoid bright-field image segmentation method provided in any embodiment of this application is executed. This application also provides a computer-readable storage medium storing computer-executable instructions for performing the above-described organoid bright-field image segmentation method. In one embodiment, the storage medium stores computer-executable instructions that are executed by one or more control processors, such as one or more processors 1020 in the computer device 1000, which can cause the one or more processors 1020 to perform the organoid bright-field image segmentation method provided in any embodiment of this application. The embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium. The above provides a detailed description of the preferred embodiments of this application, but this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and all such equivalent modifications or substitutions should fall within the scope defined by the claims of this application.

Claims

1. A method for bright-field image segmentation of organoids, characterized in that, The method includes: Obtain the original microscopic image of a high-throughput bright-field organoid containing a bounding box, and crop each organoid instance from the original microscopic image into an independent single-instance image based on the bounding box; The single instance image is segmented into multiple superpixel regions, and the skewness parameter and mean parameter are extracted based on the intensity distribution of each color channel pixel value in each superpixel region to form a skewed Gaussian embedding feature vector. Clustering is performed using the skewed Gaussian embedding feature vector as input to obtain boundary class superpixels and non-boundary class superpixels to generate an initial boundary binary image; Based on the initial boundary binary image and boundary completion algorithm, an initial mask proposal is generated; The network is trained based on the initial mask proposal and the segmentation network to output the final pixel-level segmentation mask.

2. The method according to claim 1, characterized in that, The single-instance image is segmented into multiple superpixel regions. Based on the intensity distribution of pixel values ​​in each color channel within each superpixel region, skewness and mean parameters are extracted to construct a skewed Gaussian embedding feature vector, including: The single instance image is segmented into multiple superpixel regions with similar texture and color features by using a simple linear iterative clustering algorithm. For each superpixel region, extract the intensity distribution of pixel values ​​of each color channel within the superpixel region and fit a skewed Gaussian distribution; The skewness parameter and mean parameter are obtained from the skewed Gaussian distribution to form the skewed Gaussian embedding feature vector.

3. The method according to claim 2, characterized in that, The single-instance image is segmented into superpixel regions based on a simple linear iterative clustering algorithm to obtain multiple superpixel regions with similar texture and color features, including: A five-dimensional space is established, consisting of a color space and two-dimensional coordinates, wherein the color space includes the RGB color space or the CIELAB color space; Obtain the superpixel count parameter at multiple scales; In the five-dimensional space, the single instance image is segmented into multiple superpixel regions with similar texture and color features by using a simple linear iterative clustering algorithm.

4. The method according to claim 1, characterized in that, The step of generating an initial mask proposal based on the initial boundary binary image and the boundary completion algorithm includes: The centroid coordinates of the organoid are obtained from the initial boundary binary image; According to the boundary completion algorithm, rays are emitted from the centroid coordinates of the organoid to the sampling points of the four sides of the bounding box to detect the boundary pixels in each direction. The direction where there are no boundary pixels or the boundary pixels change abnormally is completed to generate a closed boundary image. Perform region filling on the closed boundary image to generate multiple candidate masks; Multiple candidate masks are fused through voting to generate an initial mask proposal.

5. The method according to claim 4, characterized in that, The step of emitting rays from the centroid coordinates of the organoid to the sampling points on the four sides of the bounding box according to the boundary completion algorithm to detect boundary pixels in each direction, and completing the boundary pixels in directions where there are no boundary pixels or abnormal abrupt changes in boundary pixels to generate a closed boundary image includes: According to the boundary completion algorithm, rays are emitted from the centroid coordinates of the organoid to the sampling points on the four sides of the bounding box to detect boundary pixels in each direction. When the boundary pixel exists, the boundary radius of the boundary pixel is recorded as the effective boundary radius; When the boundary pixel is missing or undergoes an abnormal change, it is filled in according to the radius statistics of adjacent effective directions to obtain the effective boundary radius; A closed boundary image is generated based on the effective boundary radius.

6. The method according to claim 4, characterized in that, The process of performing region filling on the closed boundary image generates multiple candidate masks, including: Obtain the centroid coordinates of the target within the region enclosed by the closed boundary image; Using the target centroid coordinates as the seed point, a flood fill algorithm is executed to set all pixels within the closed contour of the closed boundary image as foreground values, generating multiple candidate masks. The flood fill algorithm starts from the seed point and spreads inward in a predetermined number of neighborhoods until it encounters the boundary pixels of the closed boundary image and stops spreading.

7. The method according to claim 1, characterized in that, The step of training based on the initial mask proposal and the segmentation network to output the final pixel-level segmentation mask includes: Using the initial mask proposal as a pseudo-label, the segmentation network is iteratively trained to obtain the trained segmentation network. After each training round, the predicted mask generated by the trained segmentation network is obtained, and pseudo-labels are selected or fused from the initial mask proposal and the predicted mask according to the dynamic decay ratio to generate pseudo-labels for the next training round. The process continues until the preset number of iterations is reached, at which point the final pixel-level segmentation mask is output.

8. A bright-field image segmentation device for organoids, characterized in that, The device includes: The acquisition module is used to acquire the original microscopic image of a high-throughput bright-field organoid containing a bounding box, and to crop each organoid instance from the original microscopic image into an independent single-instance image based on the bounding box. The segmentation module is used to perform superpixel segmentation on the single instance image to obtain multiple superpixel regions, and extract skewness parameters and mean parameters based on the intensity distribution of pixel values ​​of each color channel in each superpixel region to form a skewed Gaussian embedding feature vector. The clustering operation module is used to perform clustering operations with the skewed Gaussian embedding feature vector as input to obtain boundary class superpixels and non-boundary class superpixels to generate an initial boundary binary image. The generation module is used to generate an initial mask proposal based on the initial boundary binary image and the boundary completion algorithm; The output module is used to train the network based on the initial mask proposal and the segmentation network, and output the final pixel-level segmentation mask.

9. A computer device, characterized in that, include: Memory, used to store programs; A processor for executing a program stored in the memory, wherein when the processor executes the program stored in the memory, the processor is configured to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program for causing a computer to perform the method as described in any one of claims 1 to 7.

Images