METHOD AND DEVICE FOR THE BIOMETRIC ANALYSIS OF A HUMAN EYE

Abstract

A computer-implemented method for the biometric analysis of an eye is provided, based on image data acquired by optical coherence tomography. The method involves determining curve parameter values ​​from the image data using an artificial intelligence-based algorithm. The curve parameter values ​​describe at least one curve (111, 112, 121, 122, 131) that runs along at least one interface (11, 12, 21, 22, 31) of at least one ocular structure (1, 2, 3) in the image data.

Description

[0001] The present disclosure relates to a computer-implemented method for biometric analysis of an eye based on image data acquired by means of optical coherence tomography of an eye, as well as to a computer-implemented method for training an artificial intelligence-based algorithm which, after training, is configured for biometric analysis of an eye based on image data.

[0002] The present disclosure further relates to a data processing device, comprising means for executing the biometric analysis method and / or the training method. It is also conceivable to provide a computer program or computer program product, comprising instructions that, when executed by a computer, cause the computer to execute the biometric analysis method and / or the training method. It is also conceivable to provide a computer-readable medium, comprising instructions that, when executed by a computer, cause the computer to execute the biometric analysis method and / or the training method.

[0003] Furthermore, a training data set can be provided with which an artificial intelligence-based algorithm can be or is trained such that, after training according to the method for biometric analysis, it determines values ​​of curve parameters based on the image data that describe at least one curve along at least one interface of at least one eye structure of the eye. Furthermore, an algorithm or model can be provided which can be or is trained such that, after training the image data according to the method for biometric analysis, it determines values ​​of curve parameters that describe at least one curve along at least one interface of at least one eye structure of the eye.

[0004] A cataract (also known as cataract or cataractous cataract) is a condition in which the lens of a person's eye becomes cloudy. As the condition progresses, and assuming a symptomatic course, the clouding and associated loss of visual acuity, as well as sensitivity to glare, increase in the affected eye. Depending on the stage of the cataract, surgical treatment of the affected eye may be indicated. During cataract surgery, the lens of the affected eye is regularly removed, for example, after a cataract incision, and replaced with an artificial lens, optionally an intraocular lens (IOL).

[0005] In preparation for and afterward for such a surgical procedure, the affected eye is measured, and the biometric data obtained from the measurement is analyzed. This means that the affected eye is traditionally measured before or preoperatively and after or postoperatively. More specifically, in preparation for cataract surgery, biometric parameters of the eye to be operated on, such as the axial length of the eye, the central thickness of the cornea, the depth of the anterior chamber, and / or the lens thickness, are determined in order to select a suitable intraocular lens for the respective eye, optionally with regard to its refractive properties, based on the determined biometric parameters.

[0006] For example, it is important to calculate the correct IOL power based on the measured biometric parameters to minimize refractive surprise after cataract surgery. After cataract surgery, the postoperative assessment (ie, a follow-up examination) includes determining, among other things, IOL position, IOL tilt, and / or posterior capsular opacification.

[0007] Traditionally, the measurement of the eye is performed using optical measurement technology. OCT B-scan imaging (see below for details) is widely used in the ophthalmological examination or diagnosis of the cornea, anterior chamber, lens, retina, etc., e.g., using an OCT biometric system (e.g., a time-domain OCT system, a spectrum-domain OCT system, or a swept-source OCT system). Such a system typically comprises an OCT light source, an OCT detector, a processing unit (with memory), and a display device.

[0008] With a sufficiently large measurement or scanning depth, optical coherence tomography is suitable for measuring the entire axial length of the eye. More precisely, optical coherence tomography (OCT) can initially resolve a structure of the eye in a one-dimensional manner in the axial direction or along a measurement axis. For this purpose, a so-called A-scan, i.e., a single scan, is acquired or a single measurement is performed along a measurement axis. The A-scan shows the respective optical distances between the anterior corneal surface of the eye, the retina of the eye, and any other structures along the measurement axis.

[0009] In order to measure the structure of the eye two-dimensionally and thus obtain a two-dimensional image of the eye, i.e., a so-called B-scan or cross-sectional image (also called a tomogram when infrared light is used), the measurement axis can be shifted, for example, using a mirror system, and an A-scan can be performed along each of the multiple measurement axes. Furthermore, measurements can be performed along measurement axes that run obliquely or at an angle, optionally perpendicular, to the measurement axes of the B-scan in order to obtain a three-dimensional image of the eye, i.e., a so-called C-scan.

[0010] In both use cases (preoperative, i.e., determination of (accurate) biometric parameters, and postoperative, i.e., verification and / or confirmation of correction positioning), a correct and robust localization of the surfaces of the anatomical structure of the eye in the OCT scan, e.g., in the OCT B-scan, is required. Although a complete OCT scan of the eye allows the cornea, lens, and retina to be distinguished based on their spatial position and a rough location of the cornea, lens, and cornea is recognizable, a sufficiently accurate (and optionally rapid) determination of their respective spatial boundaries or interfaces is challenging.

[0011] The challenges of localizing the surfaces of the anatomical structure of the eye are optionally due to the large variability of the measurement data, which may occur due to, for example, physiological differences (e.g., different shape and / or size of the eye, different lens types (phakic, i.e., natural, or pseudophakic, i.e., artificial, lens) and / or floaters in the vitreous body), pathologies (e.g., cataract, keratoconus and / or macular degeneration), different signal characteristics (i.e., a wide spectrum of signal intensities due to different light reflection, scattering and / or absorption of the ocular tissue), measurement of one and the same eye under different conditions (e.g., regarding pupil size, examination time, pre- / post-operative evaluation), and / or noise or other disturbances (e.g., eye movements).

[0012] Problems and challenges in both use cases involving OCT scans arise, among other things, from the total time from OCT data acquisition through data processing, data analysis, and result evaluation, which should be as short as possible to utilize the time of patients, physicians, and technical assistants as efficiently as possible and without wasting time. This means that the localization of surfaces in OCT B-scans should not only be robust and precise, but also fast. Especially considering the hardware limitations of typical OCT biometer systems, postoperative solutions should also be suitable for low-performance devices.

[0013] A widely used method for analyzing OCT B-scans is segmenting them, e.g., using conventional segmentation algorithms. Conventional segmentation algorithms for OCT B-scans are usually based on stepwise image processing, which first involves feature extraction (also known as transformation) followed by semantic segmentation (also known as filtering). Raw data is input pixel by pixel and transformed into a specifically defined feature space. Semantic segmentation filters out irrelevant content from the feature space and highlights predefined object content, such as strong edges, peak distribution, minimum or maximum entropy, etc.

[0014] In OCT images for biometrics, the gradient of the pixels is often calculated as a key feature, since OCT imaging has two-armed interference on the layer structure of the human eye, with the further differentiation of individual surfaces from a gradient map often based on anatomical knowledge of the human eye. Two-armed means that a sample beam (from one measuring arm) and a reference beam (from a reference arm) interfere with beam path differences, with the interference being high when the measuring beam passes from one optical medium into another due to different refractive indices. The reconstruction (i.e. the OCT B-scan) of the interference signal therefore shows strong intensity signals and thus large gradients in these areas, particularly at the interfaces of the respective optical media in the eye.to locate a phakic lens or a pseudophakic lens between the cornea and retina, where the cornea typically has two surfaces spaced apart along the visual axis by 400–500 µm. Such conventional algorithms, based on ophthalmological expertise and clinical requirements, are used for OCT segmentation, i.e., both for OCT B-scans of the retina and for OCT B-scans of biometry.

[0015] In this context, US 10,123,689 B2 describes a method for editing a retinal layer boundary segmentation, comprising receiving a set of structural OCT B-scans and segmenting the retinal layer boundaries of the structural OCT B-scans.Segmenting includes receiving a set of relative retinal layer thickness data, enhancing the intensity transition in the structural OCT B-scan images, selecting a first B-scan image from the set of structural OCT B-scans, selecting a first retinal layer boundary in the first B-scan image, applying a directional graph search to the first retinal layer boundary, thereby generating a first segmented retinal layer boundary, selecting a second retinal layer boundary on the first B-scan image, restricting a search area of ​​the second retinal layer boundary using the first segmented retinal layer boundary and the set of relative retinal layer thickness data, and applying a directional graph search to the second retinal layer boundary. This generates a second segmented retinal layer boundary.The method further comprises checking the set of first and second segmented retinal layer boundaries, selecting a segmented retinal layer boundary in a B-scan, editing the segmented retinal layer boundary in a B-scan using a smart scissors tool based on a directed graph search, or a portion thereof, thereby generating a new retinal layer boundary segmentation, and propagating the new retinal layer boundary segmentation to neighboring structural OCT B-scans.

[0016] The conventional, classical image processing methods described above have a number of disadvantages, e.g., with regard to their robustness, a dependence on assumptions, a partial lack of precision in borderline cases, and their lack of transferability to similar use cases. In detail, this means that conventional segmentation approaches sometimes reach their limits when it comes to the (correct) detection of IOL surfaces with low signal intensity. It can also be observed that conventional segmentation approaches sometimes confuse surfaces with high signal intensity and OCT detection errors. In addition, the expertise or assumptions determine, among other things, the selection of the feature space and the step-by-step feature filtering. If the expertise is not representative or the assumptions are incorrect, the resulting segmentation accuracy is limited.Furthermore, algorithm design is often guided by expertise derived from academic statistics on human health or from clinical experience with a large number of diseased patients. Therefore, such conventional classical algorithms typically reach their limits when segmenting a healthy eye with unusual characteristics or a diseased eye with unpredictable diseases / disorders. The same applies to borderline cases, for example, when two objects are close to each other or even adjacent in an OCT scan, such as floaters in the vitreous and / or capsular bag directly in front of and / or behind a phakic lens or IOL.

[0017] Furthermore, since the implementation of feature extraction and segmentation techniques must be adapted to each application case, conventional, classical image processing involves a high technical effort and cannot be easily transferred to the next application case with different application conditions and prior knowledge.

[0018] In contrast to the conventional, classical image processing described above, deep learning techniques aim to implement both steps—feature extraction and semantic segmentation—in a single algorithm, whose parameters are all jointly estimated or learned by optimizing a (optionally single) loss function. Examples of deep learning algorithms include artificial neural networks (ANNs), optionally convolutional neural networks (CNNs), and, possibly in combination with convolutional neural networks or alternatively, transformers.

[0019] In traditional OCT segmentation methods, domain knowledge is often explicitly incorporated into the algorithm and the step-by-step algorithmic analysis. In contrast, in machine learning (optional in deep learning algorithms), domain knowledge is mainly incorporated implicitly through marked or labeled training data (also known as supervised machine learning). The training data can comprise datasets, each of which contains raw data (e.g., pixel-by-pixel B-scan images) and a result (through annotations) expected by the artificial intelligence based on the input of the raw data. Through this relationship between input data and output data, the corresponding domain knowledge is implicitly incorporated during the training process of the artificial intelligence-based algorithm. The trained neural network can then react to unknown data (e.g.,OCT B-scans of previously unknown patients).

[0020] The segmentation of medical image data is a central problem in medical image analysis. Segmentation is particularly important for computer-assisted medical diagnostics and therapy, as it forms the basis for further analysis, measurement, and 3D visualization of medical image objects. Segmentation refers to a process for creating regions with related content by combining neighboring pixels or voxels according to a specific homogeneity criterion. The result of segmentation is usually a so-called segmentation map, in which (assuming binary segmentation) the pixels of the segmented image data are assigned to a foreground or background.

[0021] In this context, US 10,198,832 B2 describes a segmentation and subsequent classification of OCT scans using neural networks for tissue classification and disease screening. The method described in US 10,198,832 B2 is particularly advantageous for tissue classification in which several adjacent tissue layers in the human retina must be distinguished, because pixel-by-pixel segmentation may not be sufficient to distinguish between the tissue layers. According to US 10,198,832 B2, the various tissue types of the eye are segmented from one another in a segmentation map, i.e., the pixels of the image data are assigned to different tissue types.

[0022] However, US 10,198,832 B2 does not provide a method by means of which interfaces of eye structures and / or an IOL, optionally spaced from one another along the eye axis, can be segmented in OCT-B scans for biometric applications, such as the preoperative and / or postoperative analysis of an eye during cataract surgery described above.

[0023] The not yet published international patent application PCT / EP2023 / 054412 describes a two-step approach.

[0024] In a first step, image data, e.g., OCT B-scans, are segmented, and a probability map (also known as a segmentation map) is created based on the image data using an artificial intelligence-based algorithm that includes a convolutional neural network (CNN). For each pixel of the image data, the probability map contains a probability value or probability vector indicating whether the respective pixel can be assigned to an interface(s) of an eye structure. Each pixel can be assigned to the foreground (i.e., interface) or background (no interface).

[0025] In a second step, parameterized interface descriptions can be adapted to this probability map. For this purpose, position data of the pixels assigned to the interface(s) of the eye structure and at least one curve based on the position data can be determined, wherein the at least one curve indicates a position or positions of the interface(s) of the eye structure. The determination of the at least one curve can be performed by curve fitting (polynomial curve fitting), whereby a given mathematical model function is optimally adapted to data points, here the position data of the pixels assigned to the interface of the respective eye structure. This allows, for example, five curves to be determined for the five interfaces of the eye structure(s) (the posterior and anterior interfaces of the cornea and the lens, as well as the interface of the retina).

[0026] The approach presented in PCT / EP2023 / 054412 leads to promising results and offers advantages over conventional image processing methods in terms of accuracy of method results, robustness, scalability and transferability together with low implementation effort.

[0027] Against the background of this prior art, one object of the present disclosure can be seen in providing a device and a method each suitable for at least enriching the prior art described above. Optionally, a sufficiently robust and precise extraction of position parameters of interfaces of spatially separated eye structures (e.g., cornea, lens, and / or retina) and / or an IOL from an OCT B-scan of the eye should be enabled for biometric applications.

[0028] The problem is solved by the features of the independent claim. The subclaims contain preferred developments of the invention.

[0029] The task is then solved by a computer-implemented method for the biometric analysis of an (optionally human) eye based on image data acquired from the eye using optical coherence tomography.

[0030] The method comprises determining values ​​of curve parameters based on the image data using an artificial intelligence-based algorithm. The values ​​of the curve parameters describe (and / or define, represent) at least one curve that runs in the image data along at least one boundary surface of at least one eye structure. The values ​​of the curve parameters can be determined directly by the artificial intelligence-based algorithm. In other words, the image data can represent input data for the algorithm, and the values ​​of the curve parameters can represent or correspond to output data of the algorithm.

[0031] Biometrics is a science that deals with measurements on living beings and the measurement and evaluation methods required for this purpose. The present disclosure can be located within this technical field. More specifically, the present disclosure relates to a method for analyzing image data or measurement data that were or are acquired from a human eye using an optical measurement method, namely optical coherence tomography.

[0032] Computer-implemented means that at least one of the, optionally all, steps of the method are carried out by a data processing system or a computer.

[0033] The at least one curve can be (respectively) limited or (respectively) have a start and an end point. The at least one curve can therefore also be referred to as part of a curve or partial curve. However, this is not necessarily the case, i.e., the prediction of the start and / or end point can optionally be made in addition to the prediction of the curve parameters. Furthermore, in the case of a two-dimensional representation, e.g., an image, the start and / or end point can also result from the outer boundaries of this representation. A prediction of the start and / or end point can then be omitted.

[0034] The values ​​of the curve parameters can describe several curves, each of which runs along one of several interfaces of the at least one eye structure in the image data. The multiple eye structures can be arranged spatially separated from one another. The interfaces of the eye structures are thus also spatially (far) apart from one another.

[0035] In the present application of biometrics, one goal is to determine the positions and curvatures (or courses) of interfaces of the eye structure(s), whereby the eye structures can be spatially (far) apart from each other and arranged along the measurement axis as follows: cornea, lens, retina.

[0036] Accordingly, the at least one eye structure or structures are (at least) a lens, optionally a natural lens and / or an artificial lens, and a cornea and / or a retina of the eye.

[0037] It is conceivable that the position (or course) of a posterior and / or anterior interface of the cornea and / or the body's own lens and / or an anterior interface of the retina may be determined. This can optionally be done preoperatively and / or postoperatively.

[0038] The artificial lens can be an aphakic intraocular lens (IOL). The aphakic IOL is implanted after removal of the natural lens (aphakia), e.g., during cataract surgery. An intraocular lens is constructed from a central optical lens and a peripherally attached haptic that fixes the central optical lens in the eye. Using this method, for example, the position and curvature (or course) of a posterior interface and / or anterior interface of the IOL can be determined. In addition to the aphakic IOL, a phakic IOL can also be used, which is inserted in addition to the natural lens. In this case, the method can be used to determine the position (or course) of a posterior interface and / or anterior interface of the phakic IOL and / or a posterior interface and / or anterior interface of the body's own or natural lens.

[0039] An interface can be understood as the transition from one eye structure to a neighboring eye structure or to the surrounding area. The interface can, for example, correspond to a surface of an eye structure. In the cornea, this can be the posterior transition to the eye chamber and the anterior transition to the surrounding area, since the cornea borders the eye on the outside. In the lens, optionally the natural lens or aphakic IOL, this can be the posterior transition to the vitreous body and anterior to the eye chamber (with a phakic IOL, this can be the posterior and anterior transition to the eye chamber). In the retina, this can be the anterior transition to the vitreous body. The interface of the retina can optionally be the interface of the retinal pigment epithelium, i.e. its transition to the vitreous body. The structure of the eye is assumed in which the eye chamber is located between the cornea and the vitreous body.The retina is located adjacent to the vitreous body. Anteroposteriorly, the two corneal interfaces are located, followed by the two lens interfaces, followed in turn by the retinal interface.

[0040] The image data acquired by means of optical coherence tomography (OCT) can be acquired by means of an (optionally single) OCT biometric scan along an entire axial length of the eye, comprising the cornea, the (at least one) lens and the retina, or along a part of the entire axial length of the eye, comprising the lens and the cornea and / or the retina.

[0041] This means that the procedure can use (at least) an OCT biometric scan or OCT image data acquired using optical coherence tomography (described above) essentially along the entire axial length of an (optionally human) eye. The entire axial length can be understood as a distance from the cornea through the lens and vitreous body to the retina.

[0042] It is also conceivable that an OCT biometric scan is used which was acquired exclusively or only along a part of the entire length of the eye, for example only along the part which concerns an anterior chamber of the eye, ie comprising the lens and the cornea, or the posterior part of the eye, ie comprising the lens and the retina.

[0043] In general, however, it should be noted that the application of biometrics, in which eye structures that are relatively far apart from one another are detected and optionally their distances to one another are measured, must be distinguished from the application of diagnostics, in which tissue layers that are very closely spaced are to be detected or segmented.

[0044] In other words, in the field of (retinal) diagnostics, for example, different tissue layers of the retina and thus one and the same eye structure are analyzed, whereas in biometrics different eye structures (e.g. lens and cornea) are recognized and optionally their distance from each other is measured.

[0045] Aside from these dimensional differences between OCT biometric scans and OCT diagnostic scans, which result in significantly more background in the OCT image data in the former, another difference between the two scans is that OCT biometric scans usually also aim to detect the lens, which differs in its appearance in the OCT image data from other tissue layers, such as the retina and cornea. Therefore, an artificial intelligence-based model faces fundamentally different challenges when analyzing OCT biometric scans than when analyzing OCT diagnostic scans.

[0046] The method is therefore not intended to distinguish between structures that are close to one another or adjacent to one another by determining values ​​of curve parameters, but rather to determine the boundaries (optionally anterior and / or posterior, but possibly also lateral) of the individual eye structures that are relatively far apart from one another as accurately as possible or sufficiently.

[0047] The method can utilize image data acquired using optical coherence tomography (described above), essentially along the entire axial length of a human eye. The entire axial length can be understood as a distance from the cornea through the lens and vitreous body to the retina.

[0048] The image data can be an OCT-A scan (one-dimensional), an OCT-B scan (two-dimensional) and / or an OCT-C scan (three-dimensional), as described above.

[0049] An algorithm based on artificial intelligence can be understood as an algorithm that defines an artificial system that has learned relationships from examples or based on training data during a learning phase and can generalize these relationships after the learning phase has ended (so-called machine learning). The algorithm can have a statistical model based on training data. This means that in the machine learning described above, the examples are not memorized from the training data, but rather patterns and regularities are recognized in the learning or training data. This allows the algorithm to also evaluate unknown image data (so-called learning transfer or generalization).

[0050] The artificial intelligence-based algorithm can comprise at least one artificial neural network. An artificial neural network can be based on several interconnected units or nodes, called artificial neurons. A connection between the neurons can forward information (sometimes also referred to as a signal) to one or more other neurons. An artificial neuron receives input information, processes the input information, and can then output output information based on the processing of the input information to the neuron(s) connected to it. The input and output information can be a real number, whereby the output information of the respective neuron can be calculated using an optionally non-linear function as the sum of its input information. The connections can also be referred to as edges.Neurons and edges typically have a weight that is adjusted during the learning process or a training procedure using an optimized loss function. The weight increases or decreases the output information. Neurons can have a threshold, so that output information is only sent if the output information exceeds this threshold. Neurons can be grouped into layers. Different layers can perform different transformations on their inputs. The information can be sent from a first layer (the input layer) to a last layer (the output layer), possibly via intermediate layers, possibly after passing through one or more of the layers multiple times.More specifically, the neurons can be arranged in multiple layers, where neurons in one layer can optionally only be connected to neurons in the immediately preceding and immediately following layers. The layer that receives external data is the input layer. The layer that produces the final result is the output layer. There can be zero or more hidden layers in between. A single-layer network can also be used. Multiple connection patterns are possible between two layers. The two layers can be fully connected, meaning that every neuron in one layer can be connected to every neuron in the next layer. The layers can also be connected using "pooling," meaning that a group of neurons in one layer connects to a single neuron in the next layer, thereby reducing the number of neurons in the next layer.Networks that only have such connections between layers form a directed acyclic graph and are called feedforward networks. Alternatively, networks that allow connections between neurons in the same or previous layers are called recurrent networks.

[0051] It is conceivable that the artificial neural network has or consists of a convolutional neural network (CNN). The structure of a convolutional neural network basically has one or more convolutional layers, optionally followed by a pooling layer. This unit can in principle be repeated any number of times; with sufficient repetitions, it is then referred to as deep convolutional neural networks, which fall into the field of deep learning. A convolutional layer can also be referred to as a filter. The input data (in this case the image data) is usually available as a two- or three-dimensional matrix (two-dimensional for a grayscale image or three-dimensional for a color image, optionally RGB). The neurons are arranged accordingly in the convolutional layer. The activity of each neuron is calculated using a discrete convolution.In this process, a (relatively small) convolution matrix (filter kernel) is moved step by step over the image data or input data. The output of a neuron in the convolutional layer is calculated as the inner product of the filter kernel with the currently underlying image section. Neighboring neurons in the convolutional layer react accordingly to overlapping areas (local environments in images or the image data). The values ​​of the kernel can be learned in the learning process described above using the loss function to be optimized. In the following step, pooling, information can be discarded. One form of pooling that can be used here is so-called max pooling, where from each, for example, 2 × 2 square (other square sizes are also conceivable) of neurons in the convolutional layer, only the activity of the most active (hence "max") neuron is retained for further calculation steps; the activity of the remaining neurons is discarded.Pooling is intended to discard less relevant information.

[0052] Fundamentally, the artificial intelligence-based algorithm should be designed or trained for tasks in the field of image processing, whereby, for example, areas within the image data (such as the individual interfaces of the eye represented in the image data) can be recognized as precisely as possible in order to then be able to determine graphical or mathematical relationships to these recognized areas (such as the values ​​of the pie parameters or the curves along the interfaces).

[0053] The artificial intelligence-based algorithm can feature the (well-known) residual neural network (ResNet), which was developed for image processing and is particularly used in this field. ResNet is a CNN architecture first published in 2015 (Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun: "Deep Residual Learning for Image Recognition", December 10, 2015, available at https: / / arxiv.org / pdf / 1512.03385.pdf). The architecture consists of several so-called residual blocks, with so-called skip connections built in, which ensure that the activation of one layer is added to the output of a later layer. This allows certain layers to be skipped, thus reducing the error rate. Furthermore, this architecture allows several model options or parameters to be specified, e.g.,a number of layers, a number of filters, filter sizes, normalization layers, skip layers, activation functions.

[0054] Furthermore, ResNet (e.g., ResNet50, where 50 indicates the number of layers of the neural network) is only one specific implementation option for the artificial intelligence-based algorithm, and the use of other models or architectures is conceivable, in addition or alternatively. These additional models could be, for example, ResNext, DenseNet, MobileNet, etc., which are also CNN architectures and can be referred to as "convolution backbones." All of these models are particularly well-known for classification tasks, but can also be converted into regression models. (Well-known) transformer architectures for image processing, such as Vision Transformer (ViT) or similar, are also applicable and are particularly suitable, for example, for capturing more extensive dependencies in image data.

[0055] The artificial intelligence-based algorithm can be designed to determine the values ​​of the curve parameters through regression, optionally directly. Regression, like classification, belongs to the field of supervised machine learning. In contrast to classification, in which inputs are classified into predetermined classes, regression predicts and outputs (real or continuous) values ​​based on the inputs, in this case the image data. For this purpose, the artificial intelligence-based algorithm can have one or more regression functions or algorithms that can be used to model the relationship between the inputs and the output and to make predictions based on this relationship. For example, the artificial intelligence-based algorithm or its neural network can have a final layer that is designed to calculate the output values, i.e.The values ​​of the curve parameters are regressed or determined by regression. This last layer can also be called the regression head. The layers before the regression head can also be called the backbone.

[0056] It is conceivable that the artificial intelligence-based algorithm comprises multiple models, each trained for different eye conditions. The different eye conditions may include, for example, phakic, aphakic, and / or pseudophakic (IOL) eyes, as well as eyes with possibly various types of (severe) cataracts with peripheral, nuclear, and / or posterior subcapsular opacities (which often cannot be properly segmented by "standard" models trained only on standard eyes and IOL types). The method may comprise (first) determining an eye condition of the eye (to be analyzed / measured), optionally based on the image data, and (subsequently) selecting a model from the models that is trained for the specific eye condition.

[0057] It is conceivable, additionally or alternatively, that the artificial intelligence-based algorithm has multiple models trained for different IOL types (e.g., for different IOL materials (e.g., acrylic, silicone, PMMA) and / or for a phakic IOL and / or a piggy bag IOL (sometimes also referred to as an add-on IOL), etc.) and / or different IOL geometries (e.g., double or plano-convex / concave, positive or negative meniscus). Phakic IOLs and piggyback IOLs are rare borderline cases where two lenses need to be segmented instead of one (i.e., phakic IOL and autologous lens, or piggyback IOL + conventional IOL). With phakic IOLs and piggyback IOLs, it is likely that the autologous lens or the conventional IOL cannot be correctly detected by standard, trained models.The method may comprise (first) determining an IOL type and / or an IOL geometry of the eye (to be analyzed / measured), optionally based on the image data, and (subsequently) selecting a model from the models that is trained for the particular IOL type and / or the particular IOL geometry.

[0058] The values ​​of the curve parameters that can be determined by the artificial intelligence-based algorithm (or the above models) can be values ​​of parameters of at least one polynomial function. The at least one polynomial function can describe or define the at least one curve that runs in the image data along at least one interface of at least one eye structure. It is also conceivable to represent the interface as a spline or polynomial line that runs along the respective interface. The latter has the advantage that local deviations from an "exact" parabola are also possible. In the case of a two-dimensional determination of the course of the interface(s), x / y coordinates, for example, can be predicted as the values ​​of the curve parameters for points on the spline.

[0059] It is conceivable that the respective curve running along one of the front boundary surfaces is a parabola and / or can be described or defined by a second-order polynomial function. Accordingly, the respective curve can be described by three values ​​of the curve parameters (or by three curve parameters), which can be determined by the artificial intelligence-based algorithm (or the above models). These three parameter values ​​can correspond to the values ​​of the parameters or coefficients of the three terms of the respective second-order polynomial function, e.g., the values ​​of the coefficients a, b, and c of the polynomial function f(x) = ax. 2 + bx +c. The anterior interfaces may include the posterior interface of the cornea, the anterior interface of the cornea, the posterior interface of the lens, and / or the anterior interface of the lens.

[0060] It is conceivable that the curve running along the retinal interface is a straight line and / or can be described or defined by a first-order polynomial function. Accordingly, the curve or straight line can be described by two values ​​of the curve parameters, which can be determined by the artificial intelligence-based algorithm (or the models above). These two parameter values ​​can correspond to the values ​​of the parameters or coefficients of the two terms of the first-order polynomial function, e.g., the values ​​of the coefficients b and c of the polynomial function f(x) = bx + c. Alternatively, it is also conceivable that the curve running along the retinal interface is a parabola and / or can be described or defined by a second-order polynomial function, analogous to the anterior interfaces.

[0061] Furthermore, it is also conceivable that the respective curves can also be described or defined by other functions or by more complex polynomial functions, e.g. polynomial trains.

[0062] The values ​​of the curve parameters of the at least one curve can define at least one superior or inferior, optionally both the superior and the inferior, end of the (at least one) curve, which is / are (each) arranged in an end region (or end point) of the at least one interface. The superior end and the inferior end of the at least one curve can be arranged (e.g., perpendicular to the measurement axis of the eye) in two opposite end regions (or end points) of the at least one interface. The superior end and the inferior end of the at least one curve can be defined such that the at least one curve runs along the entire (optionally visible in the image data) at least one interface. This thus enables, for example, not only a measurement of the distances between the various interfaces, but also the measurement of a single interface, e.g., with regard to its width or length.The determination of the start and / or end point can be achieved, for example, by predicting two additional values ​​per curve using the artificial intelligence-based algorithm, which describe, for example, the y-pixel coordinate values ​​of the start or end point on the curve (the x-pixel coordinate values ​​can be determined automatically by the curve in this case).

[0063] The method can comprise determining at least one visibility parameter as to whether the at least one boundary surface of the at least one eye structure is visible in the image data. The determination of the at least one visibility parameter can be carried out based on the image data, optionally directly, by means of the algorithm based on artificial intelligence. It is conceivable that a (single) visibility parameter is determined which indicates the number of visible boundary surfaces in the image data. Alternatively, a visibility parameter can be determined for each of the, e.g., five or seven, possible boundary surfaces, which indicates whether the respective one of the possible boundary surfaces is visible in the image data, e.g., in the form of a visibility flag (yes / no) and / or a continuous value, e.g., between 0 and 1 (e.g., 0 for not visible and 1 for visible). The (e.g.,Five) possible interfaces can include the posterior interface of the cornea, the anterior interface of the cornea, the posterior interface of the lens, the anterior interface of the lens, and the interface of the retina. By determining such visibility parameters, cases can be taken into account in which not all five expected interfaces are present or visible in the OCT scans, e.g., due to an adjusted scan depth and / or due to a very dense cataract. In these cases, the “visibility,” “confidence,” or similar can advantageously be predicted for each interface. It is also conceivable that the possible interfaces include two further interfaces, e.g., an artificial lens, for each of which a visibility parameter can be determined. This also allows special cases to be taken into account in which the eye represented in the image data has an artificial lens, e.g.,an anterior chamber IOL (implantable contact lens, ICL) or a piggyback IOL, and thus has a total of seven instead of the usual five borderline cases.

[0064] The method can comprise determining analysis image data by overlaying the image data with the at least one curve described by the determined values ​​of the curve parameters, such that the at least one curve in the analysis image data runs along the at least one boundary surface of the at least one eye structure. Optionally, the method can comprise outputting the analysis image data by means of a display device. This enables a possibility of carrying out the biometric analysis, i.e. the measurement of the eye shown in the image data (e.g., the OCT B-scan), wherein the analysis can be carried out using the image data in which the boundary surfaces are identified by the overlay using the determined curves.

[0065] The method described above offers advantages over known methods, in particular the two-stage approach described in PCT / EP2023 / 054412, e.g., with regard to throughput, memory usage, annotation effort / cost, heuristics, robustness / accuracy, and maintenance effort. This will be discussed in more detail below.

[0066] Throughput: The separate fitting step of the two-stage approach is unnecessary in the method described above, since the curve parameters describing the interfaces can be directly predicted. Furthermore, deep neural networks, for example, can be faster than segmentation models for parameter regression (depending, for example, on the size of the regression head and the size of the backbone). These two aspects together can significantly reduce computation time and increase throughput.

[0067] Memory usage: Since there's no longer any need to store a probability map, and no algorithm for fitting curves or polynomials to these probability maps needs to consume memory, overall memory usage is reduced. Furthermore, deep neural networks for parametric regression (e.g., depending on the size of the fully connected regression head) can be smaller than segmentation models. Therefore, the memory required for the procedure described above can be reduced.

[0068] Annotation effort and costs: Required annotations (tuples) are easier to create than pixel-by-pixel annotations, which are required in the two-stage approach. The required annotations can be more easily reviewed and corrected manually by human experts, for example, than pixel-by-pixel annotations.

[0069] Heuristics: The determination or calculation of curve parameters is part of the trainable or trained algorithm. Therefore, heuristic postprocessing for fitting curves or polynomials to a given pixel-wise probability map, as is the case in the two-stage approach, is no longer required.

[0070] Robustness or accuracy: The direct prediction of boundary surface curve parameters is an important aspect of interest for applications. For example, an error in the prediction of curve parameters can be directly used to improve the artificial intelligence-based algorithm or model during training. In contrast, the two-stage approach is decoupled, meaning that an error in the features calculated from the semantic segmentation result through heuristic post-processing cannot be used to train the semantic segmentation model. As a result, direct prediction using the method described above is more stable and accurate.

[0071] Maintenance effort: The procedure described above only involves one step instead of two, so that, among other things, less (maintenance) effort is required for code maintenance.

[0072] Furthermore, the disclosure also relates, either in combination with or independently of the methods described above, to a computer-implemented method for training an algorithm (or model) based on artificial intelligence. After training, the algorithm is configured for biometric analysis of an eye, wherein the image data were acquired from the eye using optical coherence tomography, the biometric analysis comprises determining values ​​of curve parameters based on the image data using the trained, artificial intelligence-based algorithm or model, and the values ​​of the curve parameters describe (and / or define, represent) at least one curve that runs in the image data along at least one interface of at least one eye structure of the eye.

[0073] The computer-implemented method can be designed for training the artificial intelligence-based algorithm described above. The above description with reference to the computer-implemented method for biometric analysis of the human eye can also apply analogously to the computer-implemented method for training the artificial intelligence-based algorithm, whereby the latter can also be referred to as a computer-implemented training method. The disclosure thus also relates to a computer-implemented method that comprises both the steps of the training method, optionally as first steps, and the steps of the computer-implemented method for biometric analysis of the human eye, optionally as second steps or steps subsequent to the training method.

[0074] It is conceivable that the artificial neural network described above, optionally the CNN, is trained or taught using at least partially supervised learning (supervised learning or semi-supervised learning). In supervised learning, a mapping from input data to target data is learned. This is done on the basis of input-target pairs, the so-called training data. In supervised learning, the corresponding outputs of the artificial intelligence-based algorithm are compared with the corresponding target data for the input data. Differences between the respective output data and the corresponding target data are evaluated using a cost function. The costs of the differences are then used to adapt the artificial intelligence-based algorithm with the aim of optimizing the costs. It is possible that the cost function is a loss function.In this case, costs can be optimized by minimizing the loss function. The loss function can be a (optionally inverted) logarithmic function (since such logarithmic functions penalize large deviations from the actual to the target more severely than small deviations).

[0075] Training data can be used to train the artificial intelligence-based algorithm. The training data comprises image data acquired from an eye using optical coherence tomography, with associated target values ​​of curve parameters. The target values ​​of the curve parameters can describe (and / or define, represent) at least one curve that runs along the at least one boundary surface of the at least one eye structure of the eye in the image data of the training data.

[0076] The image data of the training data (or the image data used for training) may have been acquired along an entire axial length of the eye, comprising the cornea, the lens and the retina, or along a part of the entire axial length of the eye, comprising the lens and the cornea and / or the retina.

[0077] The target values ​​of the curve parameters can describe at least one curve (or target curve) that runs in the image data of the training data along at least one interface of the lens, the cornea and / or the retina of the eye.

[0078] Training may involve adjusting model parameters of the artificial intelligence-based algorithm by optimizing loss functions for the values ​​of the curve parameters.

[0079] A loss function (also known as a merit function) quantifies the differences or deviations between the network's output (the prediction) and the desired output (the ground truth). Standard loss functions for semantic segmentation are cross-entropy (CE), dice (coefficient) loss, and focal loss. According to a possible optimization procedure, an output of the loss function, the so-called loss value, can be used as input. The model parameters can be iteratively adjusted during the training process based on the loss value so that it should decrease over the course of training. Common optimization functions include stochastic gradient descent (SGD), SGD with momentum, and Adam.

[0080] In this case, a (separate) loss function can be provided for each (value) of the curve parameters that are or are to be determined by the algorithm. The respective loss function can quantify the deviation between an (output) value of one of the curve parameters and the target value of the corresponding curve parameter in order to converge and minimize the loss value output by the loss function (up to a desired degree of accuracy) through iterative adjustment of the model parameters. During training of the artificial intelligence-based algorithm, several or all loss functions can be optimized together, e.g., by minimizing the, optionally weighted, sum of the loss functions. If the curve parameters, as described above, include coefficients of polynomial functions, a separate loss function can be provided for each of these coefficients.

[0081] The loss functions described above are generally suitable for any task and do not explicitly consider any specialized knowledge about or specific to the specific case, such as the expected shape of objects or eye structures whose interfaces are to be identified, an illumination process, and / or a number of eye structures and / or interfaces to be identified. Loss functions specifically adapted to the specific application can include or consider these boundary conditions or application limitations.For example, a constraint may be that a posterior corneal surface should be located to the right of the anterior corneal surface in the image data (e.g., such a fitted loss function may determine the (directional) distance between the posterior corneal surface and the anterior corneal surface, and in the case of a negative distance, return a value of 0 as the loss value, and in the case of a positive distance, return the absolute value of the distance as the loss value, so that minimizing this loss value would result in the determined posterior corneal surface not being located to the left of (i.e., in front of) the determined anterior corneal surface). A constraint may be that, for normal and small pupils, lens interfaces or interfaces of the lens should be located in the central part and not in the uppermost and lowermost parts of the B-scan, since the iris is located in these areas.A constraint may be that intensity values ​​and axial intensity gradients for foreground regions should exhibit little variation (at least for the cornea and IOL). A constraint may be that the corneal surfaces or the anterior and posterior corneal interfaces are nearly parallel and convex (larger deviations may occur in corneal diseases such as keratoconus). Furthermore, a constraint may be that in phakic eyes, the anterior lens surface / lens interface is usually convex and the posterior lens surface / interface is usually concave, usually with a steeper radius of curvature than the anterior lens surface / lens interface.

[0082] Hardware properties of the OCT system used to create the images to be analyzed can be considered when training the artificial intelligence-based algorithm. Additionally or alternatively, knowledge about the appearance of certain tissues in OCT images can also be considered, i.e., how they react to OCT imaging or illumination. For example, the cornea and IOL should vary only slightly in their pixel intensities. Such application knowledge can be explicitly considered in the loss function (e.g., by predicting class boundaries only in regions with strong variations and penalizing them in regions of low intensity variation).

[0083] Loss functions adapted to the specific application can also specifically consider other biometric parameters of the human eye for the comparison of prediction and annotation, e.g., an axial length of the human eye, an anterior chamber depth, a thickness of the corneal center, an IOL thickness, a corneal curvature, etc.

[0084] Additionally or alternatively, it may be considered that the at least one eye structure or structures comprise spatially separated anterior and posterior borders of the cornea, anterior and posterior borders of the phakic lens, anterior and posterior borders of the pseudophakic lens and a single retinal layer.

[0085] To train the artificial intelligence-based algorithm, image data acquired by optical coherence tomography along an entire axial length of the eye, comprising the cornea, the lens and the retina, or along a part of the entire axial length of the eye, comprising the lens and the cornea and / or the retina, can be used.

[0086] During model training, the images or image data (the training data) can be slightly varied, optionally slightly randomly, to artificially increase the available training data. This variation is also referred to as augmentation. Common methods for this include random contrast variation and / or intensity variation.

[0087] Furthermore, a data processing device is provided, comprising means for carrying out one or both of the methods described above (i.e. the method for biometric analysis and / or the method for training). The device can be or comprise a computing device, optionally a computer. The data processing device can be part of a device designed to record image data of a human eye by means of optical coherence tomography. The data processing device can be a distributed data processing system. The data processing device can be arranged entirely or partially remotely from the device designed to record the image data of the human eye by means of optical coherence tomography.The two devices can be connected wirelessly and / or wired, for example, via the internet and / or intranet, for data exchange. It is conceivable that the data processing device represents a component of a cloud.

[0088] Furthermore, the disclosure also relates, either in combination with or independently of the methods and device described above, to a training data set for training an algorithm based on artificial intelligence, which is designed to determine values ​​of curve parameters based on the image data by means of an artificial intelligence-based algorithm for the biometric analysis of a (human) eye based on image data recorded from the eye by means of optical coherence tomography, wherein the values ​​of the curve parameters describe (and / or define or represent) at least one curve which runs in the image data along at least one interface of at least one eye structure of the eye.

[0089] Furthermore, the disclosure also relates, either in combination with or independently of the methods and the training data set and the device described above, to an algorithm (or a model) comprising an artificial neural network, wherein the algorithm (or the model) is designed with the aid of the artificial neural network to determine values ​​of curve parameters based on the image data by means of an algorithm based on artificial intelligence for the biometric analysis of a human eye based on image data recorded from the eye by means of optical coherence tomography, wherein the values ​​of the curve parameters describe (and / or define, represent) at least one curve which runs in the image data along at least one interface of at least one eye structure of the eye.

[0090] Furthermore, the disclosure also relates, either in combination with or independently of the methods and training data set and apparatus and algorithm described above, to a computer program or a computer program product comprising instructions which, when executed by a computer, cause the computer to carry out one or both of the methods described above (i.e., the method for biometric analysis and / or the method for training).

[0091] Furthermore, the disclosure also relates, either in combination with or independently of the above-described methods and training data set and apparatus and algorithm and computer program, to a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to perform one or both of the above-described methods (i.e., the biometric analysis method and / or the training method).

[0092] It is conceivable that the computer-readable medium is a physical storage medium, such as a USB stick. However, it is also conceivable that the commands are provided wirelessly, e.g., via the Internet.

[0093] Optionally, it is conceivable that regular updates of the computer program are provided, e.g., via the computer-readable storage medium. It is also conceivable that the device described above is configured to send recorded image data segmented by the computer program and the associated raw data to a predetermined recipient, e.g., the manufacturer of the device, so that the recipient can use this information for the updates. This offers the advantage that the updates can be subjected to certification.

[0094] What has been described above with reference to the methods also applies analogously to the device, the training data set, the algorithm, the computer program and the computer-readable medium and vice versa.

[0095] The features and embodiments mentioned above and explained below are not only to be regarded as disclosed in the respective explicitly mentioned combinations, but are also encompassed by the disclosure content in other technically meaningful combinations and embodiments.

[0096] Further optional details and optional advantages of the disclosure will now be explained in more detail by means of the following optional examples and optional embodiments of the disclosure with reference to the figure.

[0097] Below is an embodiment with reference to Fig. 1 and Fig. 2 described. Fig. 1 schematically shows a flow diagram of a computer-implemented method for the biometric analysis of a human eye with a preceding training method for an algorithm based on artificial intelligence, which is used during the method for analyzing image data of the eye, and Fig. Figure 2 shows a schematic representation of an OCT B-scan together with the visualization of the associated annotations / labels, which are used as part of a training dataset in the training procedure from Fig. 1 can be used.

[0098] In Fig. 1 shows a flow diagram of the computer-implemented method according to the disclosure, which can be roughly divided into two parts. Fig. 1 Part of the method marked with TV corresponds to a sequence of a method for training an algorithm based on artificial intelligence (hereinafter referred to as training method) and the Fig. 1 The part of the method marked AV relates to a method for the biometric analysis of a (human) eye based on image data, wherein at least one position and curvature (or a profile) of at least one boundary surface of at least one eye structure of an eye represented in image data is determined by means of the algorithm trained during the training method (hereinafter referred to as analysis method).

[0099] The goal of the training procedure is for the artificial intelligence-based algorithm, after completion of the training procedure, to be able to evaluate image data acquired using optical coherence tomography in such a way that the algorithm, optionally by regression, can determine values ​​of curve parameters that describe (or define and / or represent) at least one curve. The at least one curve runs as precisely as possible along at least one interface of at least one eye structure of the eye and thus specifies the position and curvature (or course) of the at least one interface as precisely as possible. In this way, several curves can be determined that specify the positions of interfaces of predetermined eye structures, in this case a retina, an artificial and / or natural lens, and a cornea of ​​the (human) eye, as precisely as possible.

[0100] The artificial intelligence-based algorithm can comprise the ResNet50 (described above) (i.e., a ResNet comprising a 50-layer neural network), which can be described as an architecture comprising a CNN or based on a CNN for image processing or image coding. Image coding can be performed without resizing. Feature representations can be fed into a series of fully connected layers of the CNN with non-linear activations. The final layer of the algorithm (or model) regresses the features or inputs on the corresponding output parameters, i.e., the values ​​of curve parameters to be determined by the algorithm. This final layer can also be referred to as the regression head.

[0101] During the training process, the target data and / or output data can be scaled so that the mean value is a specified value (optionally zero) and the standard deviation is a specified value (optionally one) to enable robust training. Furthermore, the respective mean square error for each output parameter for the regression can be minimized during the training process, optionally weighted by adjustable relative importance weights for the output parameters. Mean absolute errors can also be used. It is also conceivable to use the Minkowski loss function (which can correspond to the square error for q = 2 and the absolute error for q = 1). The Huber loss function is also conceivable. Furthermore, it is conceivable to minimize the integrals between the ground truth polynomials and the predicted polynomials.The latter offers the advantage that the deviations are not considered separately for each parameter (and then, for example, summed up), but rather a deviation between the entire set of predicted parameter values ​​and the set of target parameter values ​​is determined. Optimization of the cost function, for example, minimizing the loss function, can be performed with an Adam optimizer. The learning rate can be adjusted (scheduled).

[0102] The training procedure essentially consists of two steps S1 and S2.

[0103] In a first step S1 of the training process, a training data set is created to train the artificial intelligence-based algorithm.

[0104] For this purpose, image data acquired using optical coherence tomography, for example, from previously conducted clinical studies, are used. The image data were acquired along the entire axial length of the eye, comprising the cornea 1, the lens 2, and the retina 3, or along a portion of the entire axial length of the eye, comprising the lens 2 and the cornea 1 and / or the retina 3. The image data can be OCT B-scans, although OCT C-scans or OCT A-scans are also conceivable as image data or training data.

[0105] It is advantageous that this image data is representative of the time of inference, i.e., the analysis process during the analysis or evaluation of image data using the trained algorithm. Therefore, it is advantageous that all image data have been acquired and created by the same or an identically constructed OCT system, e.g., an OCT biometric system, if possible, cover as many or all expected image variations as possible (e.g., contrast, blur, noise, etc.), and have, for example, the same pixel resolution. It is advantageous if the expected variations in the appearance of the eye, particularly of the cornea, lens, and retina, are covered in the image data, e.g., different lens thicknesses, different eye lengths, different lens types (e.g., crystalline or artificial), etc.

[0106] These image data are annotated or labeled by adding target values ​​of curve parameters to the image data. These target values ​​of the curve parameters describe curves 111, 112, 121, 122, 131, which in the image data of the training data each run along a boundary surface 11, 12, 21, 22, 31 of the respective eye structure 1, 2, 3. In other words, in these image data, the positions or courses of the boundary surfaces 11, 12, 21, 22, 31 are specified or marked by the added target values ​​of the curve parameters.

[0107] An example of such image data, in which a cornea 1, a lens 2 and a retina 3 of an eye are visible, is shown in Fig. 2 shown.

[0108] This shows Fig. 2 above a section of an OCT-B scan, showing the cornea 1, the lens 2 and the retina 3 with their interfaces, namely the anterior interface 11 and the posterior interface 12 of the cornea 1, the anterior interface 21 and the posterior interface 22 of the lens 2 and the interface 31 of the retina 3.

[0109] This OCT B-scan can be annotated or labeled accordingly for training by specifying target values ​​of curve parameters. Fig. 2 below are target values ​​of the curve parameters of curves 111, 112, 121, 122, 131 (see Fig. 2 above). For better visibility, the Fig. 2 below only the curves 111, 112, 121, 122, 131 are shown, whereby by comparison with Fig. 2 above it can be seen that these each run along one of the interfaces 11, 12, 21, 22, 31. Fig. 2 thus visualizes the data that is predicted (i.e. output data) or is to be predicted (i.e. target data) by the artificial intelligence-based algorithm.

[0110] The target values ​​of the curve parameters can be or include values ​​of parameters of one or more polynomial functions, with each of the polynomial functions defining one of the curves.

[0111] It is thus conceivable that the curves of the anterior interfaces, i.e., the posterior interface 11 of the cornea 1, the anterior interface 12 of the cornea 1, the posterior interface 21 of the lens 2 and / or the anterior interface 22 of the lens 22, are each a parabola and can thus be described or defined by a second-order polynomial function. Accordingly, the curves 111, 112, 121, 122 can each be described by three target values ​​of curve parameters, which can each correspond to the values ​​of the parameters or coefficients of the three terms of the respective second-order polynomial function, e.g., the values ​​of the coefficients a, b and c of the polynomial function f(x) = ax 2 + bx +c.

[0112] Furthermore, the curve 131 running along the interface 31 of the retina 3 can be a straight line and / or be described or defined by a first-order polynomial function. Accordingly, the curve 131 or straight line can be described by two values ​​of the curve parameters, which can correspond to the values ​​of the parameters or coefficients of the two terms of the first-order polynomial function, e.g., the values ​​of the coefficients b and c of the polynomial function f(x) = bx + c.

[0113] If the image data, e.g. the OCT-B scans, have been recorded along the entire axial length of the eye, including the cornea 1, the lens 2 and the retina 3, so that all five boundary surfaces 11, 12, 21, 22, 31 are recognizable, as is the case, for example, in Fig. 2 is the case, the respective image data or OCT-B-scans can be annotated or labeled with a total of 14 target values ​​of the curve parameters (three target values ​​of the curve parameters for the interfaces 11, 12, 21, 22 and two target values ​​of the curve parameters for the interface 31).

[0114] In order to determine not only the courses and thus the curvatures of the interfaces 11, 12, 21, 22, 31, but also where the interfaces 11, 12, 21, 22, 31 begin and end in height, i.e. perpendicular to the axial length of the eye, in the image data, it is conceivable that additional curve parameters are provided in order to define at least one superior or inferior, optionally both the superior and the inferior, end of the curves, which is or are arranged in an end region of the respective interface 11, 12, 21, 22, 31. For this purpose, further target values ​​of these additional curve parameters can be taken into account for the annotation, e.g. two additional target values ​​of the additional curve parameters that define the start and end point of the respective curve. This can be done for one or more or all of the curves.

[0115] The target values ​​of the curve parameters (or the corresponding curves) can be determined manually by people, e.g. with the aid of appropriate software, based on the image data and added to the image data (or the training data). In doing so, the people can, for example, recognize the boundary surfaces 11, 12, 21, 22, 31 in the respective image data and determine polynomial functions that define curves along these boundary surfaces 11, 12, 21, 22, 31 as accurately as possible. It is also conceivable to use an existing algorithm for segmentation and curve fitting for annotation suggestions, with the results of this existing algorithm being manually adjusted. This could, for example, be done in two steps, with the results of the existing algorithm first being evaluated using yes / no answers (or correct / incorrect answers) and then the results rated as no or incorrect, orParameters must be changed manually to obtain the most accurate target values ​​of the curve parameters.

[0116] The training dataset is created using a large number of OCT B-scans annotated or labeled as described above. To increase the size of the training dataset, image data or OCT B-scans can be slightly randomly varied (i.e., augmented), e.g., by slight random contrast variations, intensity variations, slight random cropping, and / or slight random rotation. For example, a random (slight) upward or downward rotation and / or random (slight) left or right rotation and / or random upward or downward rolling and / or random left or right rolling can be randomly performed to simulate slight fluctuations in position during the respective acquisition. The random upward or downward rolling and / or random left or right rolling removes a random number of rows or columns from the top or bottom or left or right edge of the image and adds them to the bottom or right edge of the image.The random number of rows or columns can be significantly smaller than the total number of rows or columns in the image, for example, no more than 10% of the number of rows or columns. Since the extensions should match the expected variations in image acquisition, it may be advantageous to avoid random flipping (as this would flip the image data left and right or up and down), strong rotations (as this is not expected), and strong cropping (as the relevant areas should not be cut away).

[0117] The created training dataset is used in a second step (S2) of the training process to train the artificial intelligence-based algorithm. For this purpose, the individual image data or OCT B-scans of the training dataset are input into the algorithm.

[0118] The algorithm then outputs values ​​of curve parameters for, for example, an input OCT-B-scan, which describe curves that should run as accurately as possible along one of the interfaces 11, 12, 21, 22, 31 in the image data.

[0119] During the learning process, each value of the curve parameters is compared with the corresponding target value of the curve parameters, and the result of the target-actual comparison is entered into a loss function. In other words, each curve parameter has its own loss function, which is intended to minimize the deviation between a value of the (output) curve parameters and the corresponding target value of the curve parameters. If the curves can be described by polynomial functions, as described above, so that 14 target values ​​of the curve parameters are specified, provided all five boundary surfaces 11, 12, 21, 22, 31 are recognizable, then 14 loss functions can be provided. The optionally weighted sum of these loss functions is optimized during training, i.e. the algorithm orSizes of the algorithm are adjusted and the algorithm is repeatedly executed for the input OCT B-scan until the sum of the loss functions is minimized (up to a desired level of accuracy of the algorithm).

[0120] Once the desired level of accuracy of the algorithm has been achieved, the training process ends, allowing the algorithm to be used in the computer-implemented method for biometric analysis of a human eye based on image data acquired using optical coherence tomography, the so-called analysis method, described in detail below. In addition or as an alternative to achieving the desired level of accuracy, the training process can be terminated after a predetermined time budget or repetition budget has been reached. Additionally or alternatively, it is also conceivable for the training process to end when a desired level of accuracy has been achieved for at least a desired number of repetitions.

[0121] Furthermore, it is conceivable to take into account in the training process that different eye conditions, IOL types, and / or IOL geometries can, for example, have different curvatures of the respective interfaces or even a different number of interfaces. For example, eyes with implanted artificial lenses, e.g., anterior chamber IOLs (implantable contact lenses, ICLs) or piggyback IOLs, have a total of seven interfaces instead of the five described above. Therefore, the artificial intelligence-based algorithm can comprise multiple models for such different eye conditions, IOL types, and / or IOL geometries. In addition to what is described above, a preselection or separation of the training data set for the multiple models of the algorithm takes place in the first step S1, and the models are trained separately from one another in the second step S2 using the respectively assigned part of the training data set.

[0122] As in Fig. 1, the analysis method may comprise three steps S3 to S5, wherein steps S3 and S5 are optional steps.

[0123] If the artificial intelligence-based algorithm, for example, comprises several models (as described above) trained for different eye conditions, IOL types, and / or IOL geometries, an eye condition, an IOL type, and / or an IOL geometry of a (human) eye represented in image data, e.g., an OCT B-scan, can be determined in an (optional) first step S3 of the analysis method. This can be done using the OCT B-scan, which was acquired, for example, in preparation for cataract surgery. The eye condition, the IOL type, and / or the IOL geometry can also be provided as additional information, e.g., via a data interface that receives the data from a local storage and / or a cloud storage. Based on the determined eye condition, the determined IOL type, and / or the determined IOL geometry, a corresponding model can be selected from the algorithm's models.

[0124] In a second step S4 of the analysis method, the image data of the OCT B-scan, which is unknown to the trained algorithm, are input into the trained algorithm (or into the selected model of the trained algorithm), whereupon the latter determines values ​​of curve parameters as described above (e.g. directly by regression), wherein these values ​​describe at least one curve which runs in the image data along at least one of the interfaces 11, 12, 21, 22, 31 of the cornea 1, the lens 2 or the retina 3.

[0125] Are, as in Fig.2, all interfaces are recognizable, the algorithm can, for example, output values ​​of 14 curve parameters that correspond to the coefficients of the polynomial functions that define the curves along the interfaces 11, 12, 21, 22, 31. In addition, values ​​of 10 curve parameters can be output, whereby these 10 curve parameters correspond to the start and end points of the five curves representing the interfaces 11, 12, 21, 22, 31.

[0126] In an (optional) third step S5 of the analysis method, the determined values ​​of the curve parameters or the determined curves can be evaluated accordingly to determine or read the position, curvature, extent, etc. of the respective interface 11, 12, 21, 22, 31. It is also conceivable that distances between two interfaces 11, 12, 21, 22, 31 and / or widths of the cornea 1, the lens 2, and / or the retina 3 are determined.

[0127] It is also conceivable that, in the third step S5, analysis image data are first determined by overlaying the image data with the at least one curve described by the determined values ​​of the curve parameters, and the analysis image data are then output, for example, via a display device. An evaluation can thus be carried out, for example, based on the determined or displayed analysis image data.

[0128] The analysis method described above can be used, for example, preoperatively to determine biometric parameters of an eye to be operated on and postoperatively, i.e., to check the surgical success, in cataract operations, since it enables the most accurate and robust analysis possible, e.g., measurement, of the interfaces of the anatomical structure of the eye in the OCT scan, optionally in the OCT-B scan.

[0129] Both the training and analysis processes are typically computationally intensive, so specialized hardware can be used to reduce the respective computation times. For example, specialized graphics processors can be used to complete training in a reasonable time (e.g., within a few days). To reduce the computation time for the inference of the trained algorithm or trained model—i.e., the determination and output of the curve parameter values—in the analysis process, specialized inference hardware can be used, such as graphics processors and / or chipsets specialized for neural network inference.

[0130] There are also opportunities on the software side to reduce the computing times for inferences, e.g. by using model quantization (e.g. operation in int8 / 16 (i.e. (signed) integer values ​​with 8 bits or 16 bits) as opposed to the usual Float32 / 64 precision (i.e. floating-point representation with 32 or 64 bits), which can significantly accelerate the calculations, e.g. in combination with dedicated int hardware).

[0131] In addition, there are further optimization options. For example, hyperparameters of the artificial intelligence-based algorithm for determining curve parameter values, e.g., of a neural network, e.g., a CNN, and / or the learning method for training the artificial intelligence-based algorithm can be optimized in the training process to achieve the best possible performance. For example, the learning rate of the optimization method can be optimized as a hyperparameter of the learning method. For example, the number of convolutional layers of the CNN can be optimized as a hyperparameter of the artificial intelligence-based algorithm. List of reference symbols 1 cornea 11 anterior interface of the cornea 12 posterior corneal interface 2 lenses 21 anterior interface of the lens 22 posterior boundary surface of the lens 3 Retina 31 anterior interface of the retina / retinal pigment epithelium 111 Curve along the anterior interface of the cornea 112 Curve along the posterior boundary surface of the cornea 121 Curve along the anterior interface of the lens 122 Curve along the posterior boundary surface of the lens 131 Curve along the anterior interface of the retina / retinal pigment epithelium QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] US 10,123,689 B2

[0015] US 10,198,832 B2 [0021, 0022] EP 2023 / 054412 [0023, 0026, 0065] Cited non-patent literature

[0000] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun: “Deep Residual Learning for Image Recognition”, December 10, 2015, available at https: / / arxiv.org / pdf / 1512.03385.pdf

[0053]

Claims

[1] Computer-implemented method for the biometric analysis of an eye based on image data acquired from the eye by means of optical coherence tomography, characterized by that the procedure has: - determining values of curve parameters based on the image data by means of an algorithm based on artificial intelligence, wherein the values of the curve parameters describe at least one curve which runs in the image data along at least one boundary surface (11, 12, 21, 22, 31) of at least one eye structure (1, 2, 3) of the eye. [2] Computer-implemented method according to claim 1, characterized by that the artificial intelligence-based algorithm is designed to determine the values of the curve parameters by regression, optionally directly. [3] Computer-implemented method according to claim 1 or 2, characterized bythat the values of the curve parameters describe a plurality of curves which in the image data each run along one of a plurality of boundary surfaces (11, 12, 21, 22, 31) of the at least one eye structure (1, 2, 3). [4] Computer-implemented method according to one of claims 1 to 3, characterized by that the at least one eye structure (1, 2, 3) is a lens (2), optionally a natural lens and / or an artificial lens, a cornea (1) and / or a retina (3) of the eye. [5] Computer-implemented method according to one of claims 1 to 4, characterized by that the image data recorded by means of optical coherence tomography were recorded by means of an OCT biometric scan along an entire axial length of the eye, comprising the cornea (1), the lens (2) and the retina (3), or along a part of the entire axial length of the eye, comprising the lens (2) and the cornea (1) and / or the retina (3). [6] Computer-implemented method according to one of claims 1 to 5, characterized by that the values of the curve parameters of the at least one curve define at least one superior or inferior, optionally both the superior and the inferior, end of the at least one curve, which is or are arranged in an end region of the at least one interface (11, 12, 21, 22, 31). [7] Computer-implemented method according to one of claims 1 to 6, characterized by that the procedure has: - determining analysis image data by overlaying the image data with the at least one curve described by the determined values of the curve parameters, so that the at least one curve in the analysis image data runs along the at least one boundary surface (11, 12, 21, 22, 31) of the at least one eye structure (1, 2, 3), and - optionally outputting the analysis image data using a display device. [8] Computer-implemented method according to one of claims 1 to 7, characterized by that the artificial intelligence-based algorithm comprises several models trained for different eye conditions, and / or IOL types, and / or IOL geometries, the method further comprising: - determining an eye condition, an IOL type and / or an IOL geometry of the eye, optionally based on the image data, and - selecting a model from the models trained for the particular eye condition, the particular IOL type and / or the particular IOL geometry. [9] Computer-implemented method, characterized by that the method comprises training an artificial intelligence-based algorithm which, after training, is designed for biometric analysis of an eye based on image data, wherein: - the image data were recorded from the eye using optical coherence tomography, - the biometric analysis includes determining values of curve parameters based on the image data by means of the trained artificial intelligence-based algorithm, and - the values of the curve parameters describe at least one curve which runs in the image data along at least one boundary surface (11, 12, 21, 22, 31) of at least one eye structure (1, 2, 3) of the eye. [10] Computer-implemented method according to claim 9, characterized bythat for training the algorithm based on artificial intelligence, training data is used which comprises image data recorded from an eye by means of optical coherence tomography with associated target values of curve parameters, wherein the target values of the curve parameters describe at least one curve (111, 112, 121, 122, 131) which runs in the image data of the training data along the at least one interface (11, 12, 21, 22, 31) of the at least one eye structure (1, 2, 3) of the eye. [11] Computer-implemented method according to claim 10, characterized by that the image data of the training data were recorded along an entire axial length of the eye, comprising the cornea (1), the lens (2) and the retina (3), or along a part of the entire axial length of the eye, comprising the lens (2) and the cornea (1) and / or the retina (3). [12] Computer-implemented method according to claim 11, characterized bythat the target values of the curve parameters describe at least one curve (111, 112, 121, 122, 131) which runs in the image data of the training data along the at least one interface (11, 12, 21, 22, 31) of the lens (2), the cornea (1) and / or the retina (3) of the eye. [13] Computer-implemented method according to one of claims 9 to 12, characterized by that training involves adjusting model parameters of the artificial intelligence-based algorithm by optimizing loss functions for the values of the curve parameters. [14] A data processing device comprising means for carrying out the method according to any one of claims 1 to 13. [15] Computer program and / or computer-readable medium comprising instructions which, when the program and / or the computer-readable medium is executed by a computer, cause the computer to carry out the method according to one of claims 1 to 13.

Images