Classification of choroidal neovascularization

Abstract

A method for identifying a type of choroidal neovascularization in the retina of a human eye is disclosed, comprising the steps of receiving optical coherence tomography data of the eye in a processor (S1), generating volume segments of the eye in the processor (S2) using the optical coherence tomography data and a neural network (3), and identifying a type of choroidal neovascularization in the eye using the volume segments (S3), the type of choroidal neovascularization including one or more of classic choroidal neovascularization (41) and occult choroidal neovascularization (42).

Description

[Technical field]

[0001] The present invention relates to a method and device for identifying types of choroidal neovascularization in a human eye. In particular, the present invention relates to a computer-implemented method, a computer, and a computer program product for identifying types of choroidal neovascularization in a human eye. [Background technology]

[0002] The eye contains many different types of tissues and fluids. The posterior (back portion) of the eye can be roughly divided into the following layers: the neurosensory retina, which contains second and tertiary neurons and photoreceptors, the retinal pigment epithelium layer, which is just beyond the neurosensory retina and is a layer of pigmented cells that nourishes the retinal visual cells, and the choroid layer, which is the vascular layer of the eye and contains blood vessels and connective tissue.

[0003] These layers can, of course, be further subdivided. Choroidal neovascularization (CNV) is the creation of new blood vessels in the choroid layer. These new vessels can give rise to lesions. These lesions can contain non-vascular components such as fibrotic tissue, exudative hemorrhage, pigmentation, or other features that can blur the boundaries of CNV.

[0004] Choroidal neovascularization can be divided into two types: classic and occult. Classic CNV lesions usually penetrate the retinal pigment epithelium and are therefore located anterior (in front of) the RPE. Occult lesions, on the other hand, are located below the RPE (sub-RPE). Classic CNV is much less common than occult or mixed type forms (mixed is where both classic and occult lesions are present). The common definition of CNV types used today is based on macular photocoagulation testing (MPS). To determine which type of choroidal neovascularization is present, invasive fluorescein angiography (FA) is used, where a fluorescent dye is added to the patient's circulation either intravenously in intravenous fluorescein angiography or orally in oral fluorescein angiography. An angiogram is then obtained by illuminating the retina with blue light and photographing the fluorescent green light emitted by the dye. Classic CNV is defined as bright hyperfluorescence that is clearly visible and well-demarcated in the early phase of the angiogram, with increased leakage in the later phase of the angiogram causing blurring (leakage) at the margins of the hyperfluorescence. Occult CNV is determined by irregular elevation of the RPE with stippled or granular hyperfluorescence after 1-2 minutes. As time progresses, leakage from the margins may or may not be seen. Fluorescein angiography is risky because the use of fluorescein dye can cause adverse reactions ranging from nausea to vomiting, rash, acute hypotension, and more seriously, anaphylaxis and related anaphylactoid reactions that can lead to cardiac arrest.

[0005] Classic CNV can occur in exudative age-related macular degeneration (ARMD), but can also occur secondary to other conditions, such as ocular histoplasmosis syndrome, pathologic myopia, choroidal rupture, angioid streaks, or idiopathic causes.

[0006] Optical coherence tomography (OCT) provides a method to create optical cross sections of the eye, which is non-invasive and high resolution. Currently, OCT is used to evaluate patients with ARMD. This imaging helps doctors and clinicians evaluate the tomographic sequelae of neovascular membranes and monitor treatment response. The type of light used in OCT penetrates deep into tissues, allowing tissue and fluid layers to be examined even with high reflectance. In OCT, typically many one-dimensional scans (A-scans) are made at various depths within the retina. These A-scans can then be combined to form a two-dimensional cross section of the eye in a so-called B-scan. The B-scans can then be combined to form a three-dimensional image of the eye, also called a C-scan. Advanced OCT methods, such as spectral domain OCT (SD-OCT), have reduced the acquisition time of OCT images and, together with techniques to track and compensate for eye movements and blinks, have also improved the accuracy of OCT images.

[0007] WO2016 / 154485 describes methods for detecting, visualizing, and measuring the extent of retinal neovascularization, including receiving a set of cross-sectional fundus angiograms, segmenting the set of cross-sectional fundus angiograms into layers, and generating an en face inner retinal angiogram and an en face vitreous fundus angiogram.

[0008] Abhijit Guha Roy, et al., "ReLayNet: retinal layer and fluid segmentation of macular optical coherence tomography using fully convolutional networks," Biomed. Opt. Express 8, 3627-3642 (2017), discloses a method to segment 2D retinal layers and fluid in retinal OCT data. The disclosed method uses a neural network to assign each pixel of the OCT B-scan to one of 10 classes. Summary of the Invention

[0009] It is an object of the present invention to provide a method and device for identifying types of choroidal neovascularization in a human eye. In particular, the present invention relates to a computer-implemented method, a computer, and a computer program product for identifying types of choroidal neovascularization in a human eye.

[0010] According to the invention, these objects are achieved by means of the features of the independent claims. In addition, further advantageous embodiments follow from the dependent claims and the description. According to the present invention, the above-mentioned objects are achieved in particular by a computer-implemented method for identifying a type of choroidal neovascularization in a human eye. The method includes a step of receiving optical coherence tomography data of the eye, in particular of a retinal area of ​​the eye, in a processor of the computer. In another step, the optical coherence tomography data and a neural network are used to generate a volume segment of the human eye in the processor. In another step, the processor uses the volume segment to identify a type of choroidal neovascularization in the eye. The type of choroidal neovascularization includes one or more of classic choroidal neovascularization and occult choroidal neovascularization.

[0011] In one embodiment, generating the volume segments includes the processor detecting tissue layer segments and fluid segments. In one embodiment, generating the volume segments includes detecting one or more of the following types of tissue layer segments: an internal limiting membrane segment, a retinal pigment epithelium segment, and a Bruch's membrane segment.

[0012] In one embodiment, generating the volume segments includes detecting one or more of the following types of fluid segments: intraretinal fluid segments, subretinal fluid segments, subretinal hyperreflective lesion segments, central subfield segments, and pigment epithelial detachment segments.

[0013] In one embodiment, identifying the type of choroidal neovascularization comprises giving highest weight to the following types of fluid segments: subretinal hyperreflective lesion segments, subretinal fluid segments, and pigment epithelial detachment segments.

[0014] In one embodiment, identifying the type of choroidal neovascularization comprises identifying a lesion in the eye. In one embodiment, the method further comprises calculating one or more of the following geometric parameters: height of the volume segments, width of the volume segments, and distance between the volume segments, and identifying the type of choroidal neovascularization further comprises using these geometric parameters.

[0015] In one embodiment, identifying the type of choroidal neovascularization includes using a decision tree. In one embodiment, generating the volume segments includes training a neural network using machine learning and a neural network training data set of optical coherence tomography data of a large number of eyes.

[0016] In one embodiment, identifying the type of choroidal neovascularization includes optimizing a decision tree using a decision tree training data set of optical coherence tomography data from a large number of eyes and a gradient boosting algorithm.

[0017] In one embodiment, identifying the type of choroidal neovascularization comprises optimizing a decision tree using the catboost algorithm. In one embodiment, generating the volume segments using a neural network includes using a convolutional neural network.

[0018] In one embodiment, generating the volume segments using a neural network includes using a U-Net and / or a ResNet. In one embodiment, the method further includes, in the processor, extracting A-scans and / or B-scans from the received optical coherence tomography data, and generating the volume segments further includes generating the area segments using the A-scans and / or B-scans and a neural network.

[0019] In one embodiment, receiving the optical coherence tomography data includes receiving one or more of time domain optical coherence tomography data, spectral domain optical coherence tomography data, ultrahigh speed swept source optical coherence tomography data, ultrahigh resolution optical coherence tomography data, polarization sensitive optical coherence tomography data, and adaptive optics optical coherence tomography data.

[0020] In one embodiment, identifying the type of choroidal neovascularization further includes identifying whether the eye has classic choroidal neovascularization, occult choroidal neovascularization, or classic and occult choroidal neovascularization, or no choroidal neovascularization.

[0021] In addition to a computer-implemented method for identifying a type of choroidal neovascularization in a human eye, the present invention also relates to a computer for identifying a type of choroidal neovascularization in a human eye, the computer comprising a processor.

[0022] The processor is configured to receive optical coherence tomography data of the eye, specifically the retina of the eye. The processor is configured to generate a volume segment of the human eye using the optical coherence tomography data and the neural network, and to identify a type of choroidal neovascularization in the eye using the volume segment.

[0023] Types of choroidal neovascularization include one or more of classic choroidal neovascularization and occult choroidal neovascularization. In one embodiment, the processor is configured to generate volume segments by detecting one or more tissue layer segments and / or fluid segments.

[0024] In one embodiment, the processor is configured to generate volume segments by detecting one or more of the following types of tissue layer segments: an internal limiting membrane segment, a retinal pigment epithelium segment, and a Bruch's membrane segment.

[0025] In one embodiment, the processor is configured to generate volume segments by detecting one or more of the following types of fluid segments: intraretinal fluid segments, subretinal fluid segments, subretinal hyperreflective lesion segments, central subfield segments, and pigment epithelial detachment segments.

[0026] In one embodiment, the processor is configured to identify the type of choroidal neovascularization by giving the highest weight to the following types of fluid segments: subretinal hyperreflective lesion segments, subretinal fluid segments, and pigment epithelial detachment segments.

[0027] In one embodiment, the processor is configured to identify a type of choroidal neovascularization by identifying a lesion in the eye. In one embodiment, the processor is further configured to calculate geometric data including one or more of the following geometric parameters: height of a volume segment, width of a volume segment, and distance between volume segments, and to identify the type of choroidal neovascularization by using these geometric parameters.

[0028] In one embodiment, the processor is configured to identify the type of choroidal neovascularization using a decision tree. In one embodiment, the processor is configured to generate the volume segments by training a neural network using machine learning and a neural network training data set of optical coherence tomography data of a large number of eyes.

[0029] In one embodiment, the processor is configured to identify the type of choroidal neovascularization by optimizing the decision tree using a decision tree training data set of optical coherence tomography data of a large number of eyes and a gradient boosting algorithm.

[0030] In one embodiment, the processor is configured to identify the type of choroidal neovascularization by optimizing a decision tree using the catboost algorithm. In one embodiment, the processor is configured to generate the volume segments using a neural network by using a convolutional neural network.

[0031] In one embodiment, the processor is configured to generate the volume segments using a neural network by using one or more of a U-Net and / or a ResNet.

[0032] In one embodiment, the processor is configured to generate volume segments by extracting A-scans and / or B-scans from the received optical coherence tomography data and generating area segments using the A-scans and / or B-scans and a neural network.

[0033] In one embodiment, the processor is configured to receive one or more of the following types of optical coherence tomography data: time domain optical coherence tomography data, spectral domain optical coherence tomography data, ultrahigh speed swept source optical coherence tomography data, ultrahigh resolution optical coherence tomography data, polarization sensitive optical coherence tomography data, and adaptive optics optical coherence tomography data.

[0034] In one embodiment, the processor is configured to identify the type of choroidal neovascularization by identifying whether the eye has classic choroidal neovascularization, occult choroidal neovascularization, or classic and occult choroidal neovascularization, or no choroidal neovascularization.

[0035] In addition to the computer-implemented method for identifying a type of choroidal neovascularization in a human eye, and a computer with a processor configured to identify a type of choroidal neovascularization in a human eye, the present invention also relates to a computer program product and a computer for identifying a type of choroidal neovascularization in a human eye. The computer program product comprises a non-transitory computer-readable medium having computer program code stored thereon, the computer program code being configured to control a processor of a computer such that the computer executes a step of receiving in the processor optical coherence tomography data of an eye, in particular of a retina of the eye. In another step, the processor generates in the processor a volume segment of the human eye using the optical coherence tomography data and a neural network. In another step, the processor uses the volume segment to identify a type of choroidal neovascularization in the eye. The type of choroidal neovascularization includes one or more of classic choroidal neovascularization and occult choroidal neovascularization. [Brief description of the drawings]

[0036] The invention will now be explained in more detail, and by way of example only, with reference to the drawings in which: [Figure 1] FIG. 1 shows a block diagram that diagrammatically illustrates a computer with a processor and a memory. [Diagram 2] 1 shows a flow diagram depicting a sequence of steps for identifying the type of choroidal neovascularization in an eye. [Diagram 3] A block diagram showing the hierarchical structure of volume segment types is shown. [Figure 4] 1 shows a flow diagram depicting a sequence of steps for identifying the type of choroidal neovascularization in an eye. [Diagram 5] 1 shows a flow diagram representing the sequence of steps for generating an eye volume segment. [Figure 6] FIG. 1 shows a block diagram illustrating alternative steps for identifying the type of choroidal neovascularization in an eye. [Figure 7]Illustrates B-scans before and after segmentation. [Figure 8] Optical coherence tomography experimental data of two eyes featuring classic neovascularization are shown, specifically, central B-scans with pixel masks for volumetric measures, frontal projections, and thickness maps. [Figure 9] Optical coherence tomography experimental data of two eyes featuring occult neovascularization are shown, specifically, central B-scans with pixel masks for volumetry, frontal projection, and thickness maps. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0037] In Fig. 1, reference number 1 designates a computer with one or more processors 11. The computer 1 may further include various components such as a memory 12, a communication interface, and / or a user interface. The components of the computer 1 may be connected to each other through a data connection mechanism such that they can send and / or receive data.

[0038] The term data connection mechanism refers to a mechanism that facilitates data communication between two components, devices, systems, or other entities. A data connection mechanism can be a wired connection, such as a cable or a system bus. A data connection mechanism can also include wireless communication. A data connection mechanism can also include communication over a network, such as a local area network, a mobile wireless network, and / or the Internet, which in some embodiments includes an intermediary network.

[0039] Processor 11 may include other more specific processing units, such as a system on a chip (SoC), a central processing unit (CPU), and / or a reprogrammable processing unit, such as a graphical processing unit (GPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), as well as processing units specifically configured to accelerate a particular application, such as an AI (artificial intelligence) accelerator for accelerating neural networks and / or machine learning processes.

[0040] The memory 12 includes one or more volatile and / or non-volatile storage components. The storage components may be removable and / or non-removable and may be integrated in whole or in part with the processor 11. Examples of storage components include a random access memory (RAM), flash memory, a hard disk, a data memory, and / or other data stores. The memory 12 includes a non-transitory computer-readable medium having computer program code stored thereon. The computer program code controls the processor 11 to configure the computer 1 to perform one or more steps and / or functions as described herein. Depending on the embodiment, the computer program code may be compiled or non-compiled program logic and / or machine code. Thus, the computer 1 is configured to perform one or more steps and / or functions. The computer program code defines and / or is part of a particular software application. It will be appreciated by those skilled in the art that the computer program code may be distributed across multiple software applications. In one embodiment, the computer program code also provides an API-like interface so that functionality and / or data of computer 1 can be accessed remotely, such as through a client application or through a web browser.

[0041] In one embodiment, computer 1 is implemented as a server computer accessible over a network by one or more client computing devices.

[0042] FIG. 2 is a flow diagram illustrating the sequence of steps for carrying out the present invention. In step S1, the computer 1 receives optical coherence tomography (OCT) data of the eye. Specifically, the processor 11 receives the OCT data of the human eye using a data connection mechanism. In the previous step, the OCT data is generated by an OCT system using optical interferometry, as known to those skilled in the art.

[0043] In one embodiment, the OCT data includes one or more of time domain optical coherence tomography (TD-OCT) data, spectral domain optical coherence tomography data (SD-OCT), ultrahigh speed swept source optical coherence tomography data, ultrahigh resolution optical coherence tomography data, polarization sensitive optical coherence tomography data, and adaptive optics optical coherence tomography data. In a preferred embodiment, SD-OCT data is used.

[0044] The OCT data is stored in a separate server's memory or directly in memory 12 on computer 1. If the OCT data is stored in memory 12, processor 11 can receive the OCT data directly from memory 12. If the OCT data is stored on a separate server's memory, processor 11 must first retrieve the OCT data. Processor 11 uses the OCT data to generate a three-dimensional image or model of the eye or a portion of the eye. In particular, processor 11 generates a three-dimensional image or model of a section of the retina of the eye.

[0045] In step S2, a volume segment 2 is generated by the processor 11 using a neural network 3 with an input of the OCT data. The neural network 3 is stored in the memory 12. The processor 11 uses the neural network 3 to output the volume segment 2. A volume segment 2 is a region of space in the eye, specifically in the retinal area. A volume segment 2 is a region of space in a three-dimensional image or model of a cross-section of the retina of the eye generated by the processor 11 using the OCT data. Specifically, a volume segment 2 is a region of space having substantial extent in all three dimensions such that the volume segment 2 does not correspond to merely a two-dimensional slice of nominal thickness. In one embodiment, a given tissue or fluid layer of the retinal area of ​​the eye is represented in one volume segment 2 such that a surjection function exists between the tissue and fluid layers as well as the volume segment 2. Preferably, contiguous volume segments 2, i.e. volume segments 2 adjacent to each other, do not have adjacent tissue or fluid layers of the same type.

[0046] In one embodiment, each volume segment 2 corresponds to one particular type of eye tissue or fluid, specifically a tissue or fluid layer of the retinal area, and each tissue or fluid layer corresponds to one particular volume segment 2 such that a bijective function exists between the volume segment 2 and the eye layers, i.e., the tissue layers and the fluid layers.

[0047] In one embodiment, the volume segments 2 are labeled with features, which may include one or more of a number of types of volume segments 2, as described below in the description of FIG.

[0048] In one embodiment, the neural network 3 uses convolutional layers. Preferably, the neural network 3 is adapted to efficiently handle three-dimensional input data. Preferably, the neural network 3 includes U-Net and / or ResNet architectures, since these two types of neural networks 3 can be used and / or configured to perform well when generating volume segments 2 with three-dimensional data. The neural network 3 is trained to generate the volume segments 2 using machine learning and a neural network training data set of OCT data of a large number of eyes. The neural network training data set includes defined volume segments 2 of OCT data, allowing the neural network 3 to be trained, using known methods, to generate volume segments 2 of OCT data of eyes that have not been seen before. The neural network training data set includes eyes (defined by the data of each eye) with classic choroidal neovascularization 41 and occult choroidal neovascularization 42. In one embodiment, the neural network training data set further includes eyes that do not have choroidal neovascularization (healthy eyes) (as defined by the data for each eye).

[0049] In one embodiment, the volume segments 2 of the neural network training data set are further annotated with features including a type and a descriptor, and training the neural network 3 includes training the neural network 3 such that the generated volume segments 2 are labeled with one or more features. The trained neural network 2 thus assigns a feature to each generated volume segment 2. The descriptor may include geometric parameters 6 relating to the volume segment 2, as will be explained in more detail in the description of FIG. 4.

[0050] Each volume segment 2 is a three-dimensional image, model, or representation defined by a set of voxels of the three-dimensional OCT data that belong to that particular volume segment 2. Alternatively, the volume segments 2 may be defined by a mesh, such as a polygon mesh. The volume segments 2 may or may not overlap one another. As described above, the volume segments 2 may also be labeled with features.

[0051] In step S3, the processor 11 uses the generated volume segments 2 to identify types of choroidal neovascularization 40 of the eye. The types of neovascularization identified include classic choroidal neovascularization 41 and occult choroidal neovascularization 42. The processor 11 uses one or more models, algorithms, and / or functions to identify types of choroidal neovascularization 40 within the retina of the eye.

[0052] For example, using volume segments 2 to identify the type of neovascularization, rather than multiple 2D slices, allows processor 11 to improve its ability to identify the type of choroidal neovascularization. More specifically, using volume segments 2 allows processor 11 to identify the type of choroidal neovascularization with increased specificity and sensitivity. This is because the retina of the eye is a three-dimensional physical structure and is therefore naturally suited to a three-dimensional representation. In particular, three-dimensional attributes of the retinal area of ​​the eye and the various tissue and fluid layers, such as the geometric parameters of a given tissue or fluid layer or the geometric relationships between a given tissue or fluid layer and other tissue or fluid layers, are more directly represented in a set of volume segments 2 and their associated features. The more direct and natural representation of a given fluid or tissue layer as a volume segment 2 rather than as a set of regions in 2D space in a set of 2D scans is a good example of how simpler and less complex functions (especially surjective or bijective functions) can exist between the volume segment 2 (representation) and the fluid and tissue layers (reality). In particular, fewer volume segments 2 are needed to represent the retinal area of ​​the eye than the 2D area segments in a set of 2D scans.

[0053] In one embodiment, the processor 11 is further configured to use one or more models, algorithms, and / or functions to identify the absence of choroidal neovascularization 40 and thus determine whether the eye is healthy or not. Specifically, the processor 11 is configured to identify healthy eyes, eyes with classic choroidal neovascularization 41, and eyes with occult choroidal neovascularization 42.

[0054] In one embodiment, the processor 11 uses a nearest neighbor method to identify the type of choroidal neovascularization 40 for the eye in question. The nearest neighbor method is a supervised classification method that uses a labeled classification training data set. The classification training data set includes volume segments 2 of multiple eyes with known types of choroidal neovascularization 40. The nearest neighbor method checks which eye in the classification training data set has a volume segment 2 that most closely corresponds to the volume segment 2 of the eye in question and assigns the type of choroidal neovascularization 40 of the eye in the classification training data set to the eye in question.

[0055] In one embodiment, the processor 11 uses logistic regression to identify the type of choroidal neovascularization 40 in the eye of interest. Logistic regression is a statistical model for analyzing a data set that includes one or more independent variables, and the output is a dichotomous variable. The processor 11 uses logistic regression and a labeled classification training data set that includes a large number of eye volume segments 2 with known types of choroidal neovascularization 40 to generate a model that can identify the type of choroidal neovascularization 40 in the eye of interest.

[0056] In one embodiment, processor 11 uses a support vector machine to identify the type of choroidal neovascularization 40 present in the eye at issue. The support vector machine uses a classification training data set that includes a large number of eye volume segments 2 with known types of choroidal neovascularization 40 to generate an optimal hyperplane. The optimal hyperplane is used by processor 11 to identify which type of choroidal neovascularization 40 is present in the eye at issue.

[0057] In one embodiment, processor 11 uses a separate neural network to identify the type of choroidal neovascularization 40 in the eye. The neural network optimizes a set of internal weights and biases using a classification training data set that includes a large number of eye volume segments 2 with known types of choroidal neovascularization 40. This set of internal weights and biases is then used by processor 11 to identify which type of choroidal neovascularization 40 is present in the eye in question.

[0058] In one embodiment, processor 11 uses a decision tree to identify the type of choroidal neovascularization 40 in the eye, as will be explained in more detail in the discussion of FIG. In one embodiment, the processor 11 also identifies whether any choroidal neovascularization 40 is present. The processor 11 can also identify whether both classic choroidal neovascularization 41 and occult choroidal neovascularization 42 are present.

[0059] In one embodiment, the processor 11 identifies lesions in the eye using the volume segments 2. The identified lesions are used by the processor 11 to identify types of choroidal neovascularization 40.

[0060] 3 shows a block diagram of a hierarchy of types of volume segments 2. The volume segments 2 are further classified into tissue layer segments 21 and fluid segments 22. The tissue layer segments 21 are further classified into at least the following categories: inner limiting membrane segments 211, retinal pigment epithelium segments 212, and Bruch's membrane segments 232. The fluid segments 22 are further classified into at least the following categories: intraretinal fluid segments 221, subretinal fluid segments 222, subretinal hyperreflective lesion segments 223, central subfield segments 224, and pigment epithelial detachment segments 225. These listed types of volume segments 2 are not limiting and other types of volume segments 2 and hierarchies are possible.

[0061] In step S2 of FIG. 2, as explained above, the processor 11 uses the neural network 3 to generate the eye volume segments 2 using the OCT data. In addition, using the neural network, one or more of the above types of volume segments 2 are assigned by the processor 11 to each generated volume segment 2. In total, combining the different types of volume segments 2 and the possible features that each volume segment 2 can be associated with, it has been found that more than 100 features are clinically relevant. These features are either quantitative features corresponding to, for example, the type of volume segment 2, or qualitative features relating, for example, to the location of the volume segment 2 in the eye, or geometric parameters 6 of the eye (not shown). This will be explained in more detail below.

[0062] Fig. 4 shows a flow diagram representing the sequence of steps for carrying out the invention. Between steps S2 and S3, an additional step S21 of generating characteristics of the volume segments 2 is shown. In step S21, the geometric parameters 6 of the eye volume segments 2 are calculated. As explained above in the description of Fig. 1, in step S2, the processor 11 uses the neural network 3 to generate the eye volume segments 2. Then, as shown in step S21, the processor 11 calculates the geometric parameters 6. These geometric parameters 6 relate to the volume segments 2 and their size and shape, their position relative to the geometric center of the eye's retina, as well as the spatial arrangement of the volume segments 2 on the eye's retina. Specifically, the geometric parameters 6 include one or more from the following list: height 61 of a given volume segment 2, width 62 of a given volume segment 2, distance 62 between a given volume segment 2 and another volume segment 2, volume of a given volume segment 2, thickness of a given volume segment 2, a measure of how spherical a given volume element 2 is, and a measure of how smooth the surface of a given volume segment 2 is. These geometric parameters 6 are characteristics assigned to each of the volume segments 2 and are used in a subsequent step, namely, step S3, where the processor 11 identifies the type of choroidal neovascularization 40 in the eye.

[0063] Figure 5 shows a flow diagram representing a sequence of steps for carrying out the invention. In step S1', which is an alternative to step S1 already described above with respect to Figure 1, the OCT data received by the processor 11 includes A-scans 71 and / or B-scans 72. If the OCT data received by the processor includes A-scans 71, then in optional step S11, the processor 11 combines a subset of the A-scans 71 to generate a sequence of B-scans 72. The A-scans 71 are one-dimensional scans of the retina of the eye at a given depth. The A-scans are sufficiently dense both laterally and in depth to allow for the subsequent high-resolution reconstruction of both two-dimensional cross-sections (B-scans 72) and three-dimensional images (C-scans 73) of the eye. If the OCT data includes B-scans 72, the processor 11 proceeds directly to step S12.

[0064] In step S12, the processor 11 uses the neural network 3 to generate area segments of the two-dimensional B-scans 72 (received in the OCT data or generated by the processor 11). Because the B-scans 72 are two-dimensional images, they are well suited for segmentation by neural networks 3, including convolutional neural networks, specifically U-Net and / or ResNet. U-Net is a convolutional neural network developed for segmenting biomedical images, and has an architecture designed to provide accurate and highly precise results with fewer training images than traditional fully-connected convolutional neural networks. It achieves this by having an upsampling portion in addition to an erosion portion that detects small features. The upsampling portion increases the resolution of the output image and receives context information from many of the erosion layers to generate a high-resolution output. ResNet is a neural network architecture that uses shortcut or skip connections to enable efficient training of deep convolutional neural networks.

[0065] In one embodiment, tissue layers present in the B-scan 72 are segmented with a different algorithm than the fluid layers. An algorithm is used to segment the following seven tissue layers of the retina of the eye into two-dimensional tissue layer segments: inner limiting membrane, outer plexiform layer-Henle's fiber layer, boundary of myoid and ellipsoid inner segments, IS / OS junction, inner boundary retinal pigment epithelium, outer boundary retinal pigment epithelium, and Bruch's membrane. Each B-scan 72 and the two-dimensional tissue layer segments within the B-scan 72 generated by the above algorithm are labeled with the appropriate tissue layer type.

[0066] In one embodiment, the processor 11 uses the neural network 3 to segment the fluid layers present in the B-scan 72 into two-dimensional fluid segments, as described above. Specifically, intraretinal fluid, subretinal fluid, subretinal hyperreflective lesion segments, and pigment epithelial detachment layers are detected and labeled in the B-scan 72 and their respective areas defined as two-dimensional fluid segments.

[0067] The generated two-dimensional area segments of the contiguous B-scans 72 are used to generate a volume segment 2 of the retina of the eye in step S2. Specifically, the B-scans 72, including the two-dimensional fluid and / or tissue layer segments and their respective assigned features, are used by the processor 11 to generate a three-dimensional image (C-scan 73) of the retina of the eye including the volume segment 2. Each segment has an assigned feature according to the type of tissue or fluid layer segment. Each pixel of each two-dimensional B-scan 72 either belongs to a two-dimensional area segment or does not belong to any two-dimensional area segment. If a pixel of the two-dimensional B-scan 72 belongs to a two-dimensional area segment, it is assigned a label according to its type, as listed above. Each pixel of a given two-dimensional B-scan is then mapped to a voxel of the C-scan 73 having its given feature. In one embodiment, the volume segments 2 are generated such that neighboring volume segments 2 do not include adjacent fluid layer regions or tissue layer regions of the same type. Specifically, volume segments 2 are generated such that each volume segment 2 corresponds to one tissue layer segment 2 or one fluid segment 22.

[0068] FIG. 6 shows a flow diagram illustrating a sequence of steps for carrying out the present invention. In step S3′, the processor 11 identifies the type of choroidal neovascularization 40 of the eye by using the volume segments 2, the features of the volume segments 2, and the decision tree 4. The decision tree 4 is trained or optimized by using a decision tree training data set of OCT data of a large number of eyes. The decision tree training data set includes a large number of eye volume segments 2 along with the type of choroidal neovascularization 40 present in each eye. Preferably, in addition to a large number of eye volume segments 2 having choroidal neovascularization, the decision tree training data set also includes a large number of healthy eyes, i.e., eyes without any type of choroidal neovascularization, as control or reference data. Preferably, the decision tree data set includes data of at least 1000 eyes. Preferably, each eye volume segment 2 in the decision tree training data set also includes an associated assigned feature that represents the type of the volume segment along with other descriptors. These features are either categorical, placing their associated volume segments 2 into one or more of a given set of categories, or numerical, describing some geometric attribute of the volume segments 2, for example the volume or extent of the volume segments in a particular direction, as will be explained in more detail below.

[0069] Processor 11 optimizes decision tree 4 using the decision tree training data set such that decision tree 4 accurately identifies the type of choroidal neovascularization 40 as defined in the decision tree training data set. During a particular training iteration, processor 11 predicts the type of choroidal neovascularization 40 present in the eyes of the decision tree training data set, and processor 11 then checks the accuracy of the predicted type against the actual type defined in the decision tree training data set. The accuracy of the prediction is determined using a logarithmic loss function, and parameters of decision tree 4 are iteratively optimized to minimize this logarithmic loss function. Preferably, a gradient boosting function is used to optimize decision tree 4, and parameters of decision tree 4 are iteratively optimized to minimize the logarithmic loss function. Preferably, a gradient boosting algorithm is used to optimize decision tree 4. In one embodiment, an ensemble of decision trees 4 is used with the gradient boosting algorithm. When using the decision tree 4 to identify the type of choroidal neovascularization 40, the decision tree 4 can be easily visualized and captured as a flow chart-like structure with transparent and easily understandable decision rules. To validate the performance of the decision tree 4 and ensure that it is not overtrained, a separate decision tree validation data set of at least 100 eyes not yet seen during training is used. In addition, cross-validation is used to evaluate performance by combining the training and validation data sets and then randomly splitting it into a new training data set and a new validation data set. After the processor 11 trains the decision tree 4, the decision tree 4 can be further evaluated and manually tuned or simplified to make it easier to understand so that the decision tree 4 can handle edge cases and outliers. The trained decision tree 4 will identify without error whether an eye has choroidal neovascularization or not.The trained decision tree 4 allows for excellent discrimination between classic choroidal neovascularization 41 and occult choroidal neovascularization 42, with an area under a receiver operating characteristic curve (AUROC) of 0.91 with a 95% confidence interval of 0.89 to 0.94.

[0070] In one embodiment, the processor uses the catboost algorithm to optimize the decision tree 4. The catboost algorithm is advantageous because it is robust to overfitting and because many of the features discussed above are categorical, to which the catboost algorithm is particularly well suited.

[0071] In one embodiment, the decision tree 4 weights the following types of volume segments 2 higher than other types of volume segments 2: subretinal hyperreflective lesion segments, subretinal fluid segments, and pigment epithelial detachment segments. Specifically, the subretinal hyperreflective lesion segments and pigment epithelial detachment segments are weighted higher than other types of volume segments 2 because they have been shown to be the most discriminative characteristics. This means that the presence of a volume segment 2 having characteristics of one of the previously mentioned types of volume segments 2 has a greater impact than other types of volume segments 2 when determining the classification of an eye into classic choroidal neovascularization 41 or occult choroidal neovascularization 42. Additional characteristics, such as the size or volume of a given volume segment 2, are also considered by the decision tree 4.

[0072] In one embodiment, multiple decision trees 4 are used in a Random Forest model to identify types of choroidal neovascularization 40. A Random Forest model is a collection of decision trees whose results are aggregated. The Random Forest model has the advantage of reducing bias and variance errors.

[0073] Figure 7 shows an illustration of a B-scan 72 (left) and a segmented B-scan 73 (right) with the 2D fluid segments shown in grayscale. Superimposed on the two images is a diagram of the U-Net used to perform the image segmentation.

[0074] FIG. 8 shows two samples of the retina of eye 51 (top) showing a typical case of classic choroidal neovascularization 41, and eye 52 (bottom) also showing a typical case of classic choroidal neovascularization 41. On the left are B-scans 72 of each eye 51, 52 acquired by SD-OCT, showing subretinal hyperreflective foci 511, intraretinal fluid 512, and fluid segments of subretinal fluid 513. The center panel 514 is an en face projection of the B-scan, and the right panel 515 is a thickness map of each eye 51, 52.

[0075] 9 shows two samples of the retina of eye 53 (top) showing a typical case of occult choroidal neovascularization 42, and eye 54 (bottom) also showing a typical case of occult choroidal neovascularization 42. On the left is a B-scan 72 of each eye 53, 54 acquired by SD-OCT showing a segment of pigment epithelial detachment 531. The middle panel 532 is an en face projection of the B-scan and the right panel 532 is a thickness map of each eye 53, 54.

[0076] It should be noted that although the above description presents a sequence of steps in a particular order, one skilled in the art would recognize that the order of at least some of these steps may be varied without departing from the scope of the present invention.

Claims

1. 1. A computer-implemented method for identifying a type of choroidal neovascularization (40) in a human eye, comprising: receiving optical coherence tomography data of the eye in a processor (11); a step (S2) of generating, in the processor (11), a volume segment (2) of the human eye using the optical coherence tomography data and a neural network (3), the neural network (3) being configured to receive the optical coherence tomography data as an input and to output the volume segment (2) of the human eye; a step (S3) of identifying, in the processor (11), a type of choroidal neovascularization (40) in the eye using the volume segment (2), the type of choroidal neovascularization (40) comprising one or more of classic choroidal neovascularization (41) and occult choroidal neovascularization (42); 4. A computer-implemented method comprising:

2. The method of claim 1, wherein the step (S2) of generating the volume segments (2) comprises the step of detecting tissue layer segments (21) and fluid segments (22).

3. The method according to claim 1 or 2, wherein the step (S2) of generating the volume segments (2) comprises a step of detecting one or more of the following types of tissue layer segments (21): an internal limiting membrane segment (211), a retinal pigment epithelium segment (212), and a Bruch's membrane segment (213).

4. A method according to any one of claims 1 to 3, wherein the step (S2) of generating the volume segments (2) includes a step of detecting one or more of the following types of fluid segments (22): intraretinal fluid segments (221), subretinal fluid segments (222), subretinal hyperreflective lesion segments (223), central subfield segments (224), and pigment epithelial detachment segments (225).

5. 5. The method of claim 4, wherein identifying the type of choroidal neovascularization (40) comprises assigning the highest weight to the following types of fluid segments (22): subretinal hyperreflective lesion segments (223), subretinal fluid segments (222), and pigment epithelial detachment segments (225).

6. 6. The method of claim 1, wherein identifying a type of choroidal neovascularization (40) comprises identifying a lesion in the eye.

7. 7. The method according to claim 1, further comprising a step (S21) of calculating one or more of the following geometric parameters (6): a height (61) of a volume segment, a width (62) of a volume segment, and a distance (63) between volume segments, and wherein the step (S3) of identifying the type of choroidal neovascularization (40) further comprises a step of using the geometric parameters (6).

8. The method according to any one of claims 1 to 7, wherein the step (S3) of identifying the type of choroidal neovascularization (40) comprises using a decision tree (4).

9. 9. The method of claim 1, wherein generating the volume segments (S2) comprises training the neural network (3) using machine learning and a neural network training data set of optical coherence tomography data of a large number of eyes.

10. 10. The method of claim 8 or 9, wherein identifying (S3) the type of choroidal neovascularization (40) comprises optimizing the decision tree (4) using a decision tree training data set of optical coherence tomography data of a multiplicity of eyes and a gradient boosting algorithm.

11. The method according to any one of claims 1 to 10, wherein the step (S2) of generating the volume segments using the neural network (3) comprises using a convolutional neural network.

12. The method of any one of claims 1 to 11, further comprising the step (S11) of extracting, in the processor (11), A-scans (71) and / or B-scans (72) from the received optical coherence tomography data, and wherein the step (S2) of generating volume segments further comprises the step (S12) of generating area segments using the A-scans (71) and / or B-scans (72) and the neural network (3).

13. 13. The method according to claim 1, wherein the step (S3) of identifying the type of choroidal neovascularization (40) further comprises the step of identifying whether the eye has classic choroidal neovascularization (41), occult choroidal neovascularization (42), or classic choroidal neovascularization (41) and occult choroidal neovascularization (42), or no choroidal neovascularization.

14. A computer (1) for identifying (S3) a type of choroidal neovascularization (40) in a human eye, the computer (1) comprising a processor (11) configured to execute the method according to any one of claims 1 to 13.

15. A computer program product for identifying (S3) a type of choroidal neovascularization (40) in a human eye, comprising a non-transitory computer readable medium (12) having stored thereon computer program code configured to control a processor (11) of a computer (1) so that the computer (1) executes the method according to any one of claims 1 to 13.

Images