Method and apparatus for acquiring ophthalmic image data and identifying disease

Abstract

A method converts 3D data representing a patient's eye into 2D data representing a projection of the patient's eye. The 2D data can then be used to train or tune a large-scale image processing algorithm, calculate an epidemiological distribution of pixel data in the 2D data, and / or calculate a health risk score.

Description

METHOD AND APPARATUS FOR ACQUIRING OPHTHALMIC IMAGE DATA AND IDENTIFYING DISEASEFIELD OF THE INVENTION

[0001] The present disclosure relates generally to computer imaging and diagnostics and more particularly to a system and method for acquiring ophthalmic image data and identifying disease based on the ophthalmic image data.BACKGROUND

[0002] Optical coherence tomography (OCT) is a non-invasive imaging technique that uses low-coherence light to capture micrometer-resolution, three-dimensional images from within optical scattering media. OCT is widely used in ophthalmology to obtain high-resolution images of the retina, the nerve fiber layer, and the optic nerve head. OCT can provide information about the structure and function of the eye, as well as the presence and progression of various eye diseases, such as glaucoma, macular degeneration, diabetic retinopathy, and others.

[0003] Optical coherence tomography produces three-dimensional (3D) data pertaining to three-dimensional objects such as a patient’s eyes. Although 3D data is useful with respect to some analyses, two-dimensional (2D) data can be easier to work with in certain applications. A retinal nerve fiber layer algorithm is used to generate 2D images of a patient’s eye from 3D object data of that patient’s eye. Leung et al. have described a retinal texture analysis algorithm to generate 2D images from 3D object data in “Retinal Nerve Fiber Layer Optical Texture Analysis: Involvement of the Papillomacular Bundle and Papillofoveal Bundle in Early Glaucoma” by Leung CKS, Guo PY, and Lam AKN (Ophthalmology 2022 Sep; 129(9): 1043-1055. doi: 10.1016 / j.ophtha.2022.04.012. Epub 2022 Apr 22. PMID: 35469924). Leung et al. describe how axonal fiber bundles in the Retinal Nerve Fiber Layer (RNFL) can be visualized. FIG. 1, which is taken from “Diagnostic Assessment of Glaucoma and Non-glaucomatous Optic Neuropathies via Optical Texture Analysis of the Retinal Nerve Fibre Layer” by Leung, C.K.S., Lam, A.K.N., Weinreb, R.N. et al. (Nat. Biomed. Eng 6, 593-604 (2022)) shows how RNFLOptical Texture Analysis (ROTA) allows for discernment of patterns and the extent of RNFL defects in eyes with different severity of optic nerve damage. The top image of column a 102 in FIG. 1 shows that red-free RNFL photography is effective in detecting early localized RNFL defects where the contrast of RNFL scattering between the normal and abnormal RNFL is distinct. The top image of column b 104 in FIG. 1 shows how visualization of RFNL defects with red-free RNFL photography becomes less effective over the temporal macula. The top images of columns c 106 and column d 108 show how visualization of RNFL defects with red-free RNFL photograph is ineffective in eyes with extensive RNFL thinning. In contrast, ROTA as shown in the middle images of column a 102 through column d 108 uncovers the trajectorial and optical textural details of the retinal axonal fiber bundles enabling reliable detection of RNFL defects across different stages of optical neuropathy. The borders of the RNFL defects are shown in the middle images of column a 102 through column d 108 by arrows. The bottom images of column a 102 through column d 108 show the corresponding visual fields generated using the 24-2 Swedish Interactive Thresholding Algorithm (SITA) standard test. Each of the bottom images of column a 102 through column d 108 comprises a grey-scale plot on the left and a pattern-deviation probability plot on the right with the mean deviation (MD) identified at the bottom of each of column a 102 through column d 108. For each eye, one OCT image was captured with a swept-source OCT (for ROTA) and one high-quality red-free RNFL photograph was captured with a fundus camera. All eyes had glaucomatous optic neuropathy.

[0004] The ROTA algorithm (referred to as the “first ROTA algorithm” or “first algorithm”) used is> 1 and emphasizes bright structures (e.g., axonal bundles),1 and provides gamma correction for the whole image, CL is normalization to cause imagevalues fall into the sXy e [0,1] range, P^ef is per-bscan normalization for better test- retest variability, and PZ)Xy ispixel value of the OCT volume at position x,y,z (where z is along the A-scans).

[0005] Image 202 of FIG. 2, which is taken from “Diagnostic Assessment of Glaucoma and Non-glaucomatous Optic Neuropathies via Optical Texture Analysis of the Retinal Nerve Fibre Layer” by Leung, C.K.S., Lam, A.K.N., Weinreb, R.N. et al. (Nat. Biomed. Eng 6, 593-604 (2022)), shows extraction of reflected intensity value Pz,xy (e§-> B- Scan Pixel value) and image 204 of FIG. 2 shows computation of optical texture signal Sxye [0,1] Reflected intensity refers to the physical quantity that is measured using OCT. The OCT signal at a given depth (after a Fourier Transformation step) is proportional to the reflected intensity at a given depth.

[0006] The first ROTA algorithm can be used to convert 3D data to 2D data late in the data processing pipeline. What is needed is a method for converting 3D data to 2D data that can be used early in the data processing pipeline.

[0007] Ophthalmic 2D data can be analyzed in order to diagnose a patient’s condition. However, what is needed is a method for analysis of ophthalmic 2D data that produces accurate and precise data required to diagnose a patient’s condition.

[0008] OCT data alone may not be sufficient to assess the health risk of a subject, as it may not capture the subtle changes in the retinal layers that may be associated with other diseases, such as neurological diseases, systemic diseases, or cardiovascular diseases. Moreover, OCT data may vary depending on the scanning device, the scanning protocol, the operator, and the subject's eye movement, which may affect the accuracy and reliability of the analysis. Therefore, there is a need for a method of calculating a health risk score using OCT data that can overcome these limitations and provide a comprehensive and objective evaluation of the subject's health condition.SUMMARY

[0009] A method for assessing the health of a patient’ s eye includes the steps of receiving three-dimensional optical coherence tomography data of a reference eye, segmenting the three-dimensional optical coherence tomography data into a plurality of retinal layers, and converting each of the plurality of retinal layers into two-dimensional reference data by applying a transformation operation. The method can also include the step of identifying abnormalities in the patient’s eye by comparing the two-dimensional reference data to two-dimensional patient eye data. The method can also include the step of calculating an epidemiological distribution of pixel data in the two-dimensional reference data. The two-dimensional reference data can be generated as a transformation from a reflected amplitude of the three-dimensional optical coherence tomography data. The transformation can be one of ROTA, sum, average, mean, standard deviation, or coefficient of variance or any other transformation that produces the desired result. The calculating an epidemiological distribution of pixel data in the two-dimensional reference data can comprise a multi-variable analysis. The multi-variable analysis can be based on one or more of a patient gender, a pixel value of reflectance data, a patient age, a disc size, a disc morphology, a fovea-to-disc distance, or vessel patterns. A health risk score can be calculated by a machine learning model based on the identified abnormalities. The health risk score can be based on color fundus data. The health risk score can be related to at least one of neurological disease, systemic disease, or ophthalmic disease. Each of the plurality of retinal layers can be identified by patient gender and as one of a macular region, an optic disc region, a fovea region, or a vascular region.

[0010] An apparatus having a memory storing computer program instructions and a computer readable medium storing instructions for assessing the health of a patient’s eye are also described herein.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] FIG. 1 shows how a prior art RNFL Optical Texture Analysis (ROTA) allows for discernment of patterns and the extent of RNFL defects in eyes with different severity of optic nerve damage;

[0012] FIG. 2 shows extraction of a reflected intensity value (e.g., B-Scan pixel value) and computation of an optical texture signal according to prior art;

[0013] FIG. 3 shows an optical coherence tomography (OCT) processing method;

[0014] FIG. 4 shows the result of a first ROTA algorithm when applied to linear scale reflected intensity values;

[0015] FIG. 5 shows the result of the first ROTA algorithm when applied to OCT B- Scans that underwent postprocessing for display purposes;

[0016] FIG. 6 shows the result of a second algorithm used with raw reflected intensity data;

[0017] FIG. 7 shows the result of the first ROTA algorithm used with display data;

[0018] FIG. 8 shows a data flow illustrating how the second algorithm can be used with stored data;

[0019] FIG. 9 shows a data flow illustrating how the second algorithm can be used with linear scale reflected intensity data;

[0020] FIG. 10 shows an image which was generated by the second algorithm using raw reflected intensity data;

[0021] FIG. 11 shows an image which was generated by applying the second algorithm to a choroid;

[0022] FIG 12. shows an image which was generated by using data from the second algorithm on the pixelwise attenuation coefficient of an OCT scan that was calculated from the linear scale reflected intensity data;

[0023] FIG. 13 shows an image which was generated using the second algorithm applied to display data;

[0024] FIG. 14 shows an image which illustrates an application of the second algorithm;

[0025] FIG. 15 shows a variety of projections according to different embodiments;

[0026] FIG. 16 shows a flow diagram for generating reference data based on optical coherence tomography data according to one embodiment;

[0027] FIG. 17 shows how 3D optical coherence tomography data and segmentation data are used to generate reference data for various slabs / layers of a patient’s eye;

[0028] FIG. 18 shows a flow diagram for calculating a health risk score based on OCT data; and

[0029] FIG. 19 shows a high-level schematic of a computer for implementing methods, techniques, and systems described herein.DETAILED DESCRIPTION

[0030] The present disclosure is divided into three sections. The first section pertains to a method and apparatus for generating 2D images from volumetric (i.e., three- dimensional) Optical Coherence Tomography Scans. The second section pertains to generating a reference data based on the generated 2D images. The third section pertains to a method and apparatus for calculating a health risk score based on the generated 2D images.Method and Apparatus for Generating Images from Volumetric Optical Coherence Tomography Scans

[0031] Two-dimensional images are generated from volumetric (i.e., three-dimensional) Optical Coherence Tomography (OCT) scans for use in quantitative and qualitative analysis of objects, such as an eye, between specified boundaries. In one embodiment, the images are conceptually similar to en-face images but contain information that is not present in common en-face images. The generated images can be displayed in connection with conventional images to provide more information to users. Reduction of the dimensionality of a dataset (i.e., a 3D volume scan converted to a 2D image) produces images that can be used as input to tune or train large-scale image processing algorithms (e.g., machine learning models) efficiently.

[0032] In order to aid in understanding, a conventional method for generating images from volumetric optical coherence tomography scans will be described before the new method and system are described.

[0033] FIG. 3 shows an optical coherence tomography (OCT) processing method 300 that begins with spectrometer 302 being used to generate spectrometer data (e.g., 3D data). At step 304, linear scale reflected intensity datais generated based on datafrom spectrometer 302. The linear scale reflected intensity data Rz xyis used to generate log-transformed data Lzas shown at step 306. The log-transformed data of step 306 is used to generate image contrast normalization transformed data P at step 308 which is displayed as shown in image 310. In one embodiment, display data (log-scale, image contrast normalization transformed) Pz xyis stored in a vendor specific data format file. With the knowledge of processing algorithms and additional information in the vendor specific data format file, it is possible to calculate linear scale reflected intensity data Rz xyfrom the data in the vendor specific data format file. The first ROTA algorithm described in Leung is performed on the log transformed reflected intensity data Lz xyor the display data P . Leung et al state in U.S. Patent No. 10,918,275 that the pixel data is optical density data. These are logarithmically scaled intensity values (see “Optical Density Based Quantification of Total Haemoglobin Concentrations with Spectroscopic Optical Coherence Tomography” by Cuartas-Velez, C., Veenstra, C., Kruitwagen, S. et al. (Sci Rep 11, 8680 (2021)). The terms “reflectance” and “optical density” are sometimes used interchangeably in the OCT literature. For example, see “Optical Density of Subretinal Fluid in Retinal Detachment” by Ari Leshno, Adiel Barak, Anat Loewenstein, Amit Weinberg, and Meira Neudorf er (Investigative Ophthalmology & Visual Science August 2015, Vol.56, 5432-5438).

[0034] In accordance with the present disclosure, a second algorithm is used to generate images in place of the first ROTA algorithm. It should be noted that in one embodiment, the second algorithm can be performed early in the data processing pipeline, directly after the reflected intensity data is transmitted from spectrometer 302. It should further be noted that in one embodiment the images generated by the ROTA algorithm on display data or log-transformed data are visually similar to the images generated by the second algorithm on linear scale data 304.

[0035] U.S. Patent No. 10,918,275, Leung, and “Diagnostic Assessment of Glaucoma and Non-glaucomatous Optic Neuropathies via Optical Texture Analysis of the Retinal Nerve Fibre Layer” by Leung, C.K.S., Lam, A.K.N., Weinreb, R.N. et al. (Nat. Biomed. Eng 6, 593-604 (2022). https: / / doi.org / 10.1038 / s41551-021-00813-x) provide strongevidence that the first ROTA algorithm uses the display data or the log-transformed data to perform calculations.

[0036] FIGS. 4 and 5 show images of the result of first ROTA algorithm calculations using parameters described in the ‘275 patent and publications identified above. Image 400 shown in FIG. 4 was generated using the first ROTA algorithm with linear reflected intensity data where yl=7, y2=3.7 and is shown having a very low dynamic range. It is important to note that typical B-Scans are not linear scale reflected intensity values but have undergone nonlinear postprocessing for display purposes. Image 500 shown in FIG. 5 was generated using the first ROTA algorithm with display data where yl=7, y2=3.7 rather than raw reflected intensity data. The values are typically transformed into logarithmic scale and may include additional postprocessing steps (such as denoising). Image 500 appears to be almost identical to the first ROTA algorithm images from the ‘275 patent and publications identified above. The formula used to generate images 400 and 500 is

[0037] It should be noted that the a in the formula above does not influence the grayscale display since the display data are scaled to cover a full 8-bit dynamic range.

[0038] In one embodiment, a second algorithm uses linear scale reflected intensity data. The second algorithm is

[0039] where Rref is a per-bscan normalization and / is a gamma correction as is common in image processing. Image 600 of FIG. 6 shows the result of the second algorithm used with raw reflected intensity data. Image 700 of FIG. 7 shows the result of the first ROTA algorithm used with display data. Comparing image 500 and image 700 shows that the information content provided by the first ROTA algorithm shown in image500 is visually similar to the information content provided by the first ROTA algorithm shown in image 700.

[0040] FIG. 8 shows data flow 800 illustrating how the second algorithm can be used with stored data (e.g., vendor specific data format files) if the linear scale reflected intensity data is reconstructed from the stored data. In one embodiment, the stored OCT data does not store the linear scale reflected intensity data, but display data that was generated from the linear scale data via postprocessing. It is possible to reconstruct the linear scale intensity data from the raw display data. First ROTA algorithm 808 in FIG. 8 is normally used to process an vendor specific data format file that is generated based on steps 802, 804, and 806 of FIG. 8 (which correspond to steps 304, 306, and 308 shown in FIG. 3). In order to process the linear scale reflected intensity data, the vendor specific data format file must be processed in the reverse order it was created. As such, the vendor specific data format file is processed at image contrast normalization transformed data Pzstep 806. The data processed at step 806 is then reconstructed to obtain the log transformed reflected intensity data Lzstep 804. The data processed at step 804 is then converted into linear scale reflected intensity data Rzat step 802. Linear scale reflected intensity data Rzgenerated at step 802 is then used by second algorithm at step 810 to produce an image for display.

[0041] FIG. 9 shows direct data flow 900 illustrating how the second algorithm can be used with linear scale reflected intensity data thereby avoiding additional processing steps. Spectrometer 902 outputs data that is converted to linear scale reflected intensity data at step 904. Linear scale reflected intensity datais then used by the second algorithm at step 912 to generate display images. FIG. 9 also shows the data from spectrometer 902 being processed according to the first ROTA algorithm after steps 904, 906, and 908 have been completed. At step 910, the first ROTA algorithm can then be used to generate images for display. As described above, three-dimensional data representing a patient’s eye can be converted into two-dimensional data representing a projection of the patient’s eye based on a raw linear scale reflected intensity value.

[0042] FIG. 10 shows image 1000 which was generated by the second algorithm using raw reflected intensity data. It should be noted that although both the first and second algorithms can be used to analyze different layers, the first algorithm is usually limited to RFNL boundaries because the first algorithm was specifically designed to emphasize contrasts that are relevant to the diagnosis of glaucoma in the RFNL layer. The second algorithm can be used between any possible boundary layers. FIG. 11 shows image 1100 which was generated by applying the second algorithm to a section containing the choroid.

[0043] In one embodiment, the second algorithm can be applied to other data sets. For example, instead of R , other transformations of the data can be used, such as the display data or other calculations. In one embodiment, the second algorithm is applied to display data to produce a conventional en-face slab where the axonal fiber bundle is less visible compared to the reflected intensity data. FIG 12. shows an image which was generated by using data that was transformed to calculate the pixelwise attenuation coefficient, on the pixelwise attenuation coefficient of an OCT scan that was calculated from the linear scale reflected intensity data using the formula described in “Depth-resolved Modelbased Reconstruction of Attenuation Coefficients in Optical Coherence Tomography” by Vermeer KA, Mo J, Weda JJ, Lemij HG, de Boer JF (Biomed Opt Express. 2013 Dec 23;5(l):322-37. doi: 10.1364 / BOE.5.000322. PMID: 24466497; PMCID: PMC3891343) FIG. 13 shows image 1300 which was generated using the second algorithm applied to display data. Image 1300 is equivalent to an OCT slab with additional gamma correction.

[0044] In one embodiment, the second algorithm is generalized. Let a(x, y) G IRWzdenote the data points inside an A-scan at position (x, y).

[0045] A reduction function is denoted that takes a vector and produces a single scalar value:

[0046] Then the pixel-value S(x,y) in the 2D image via the generalized second algorithm is given.

[0047] These calculations can be performed in parallel. Different data for the A-Scan can also be used. The data can be raw reflected intensity values, but also display data,attenuation data, or other type of display or attenuation data derived from raw reflected values with additional calculations.

[0048] It should be noted the above equations are not limited to reductions S. The data a(x,y) could also be transformed using a transformation T: RAN->R N before applying the reduction and the processing can be still be performed in parallel. Non-local transformations of the OCT volume can be performed. For example, 3D processing of the OCT volume can be performed before applying the reduction to the 2D data. This is the one of the most general cases and encompasses the local transformation T above. However, it loses the parallelism.

[0049] In one embodiment, a summation algorithm is used. The summation algorithm is a special case of the generalized algorithm: the reduction is the sum of the Linear Scale Reflected intensity data. The first ROTA algorithm is also encompassed in this general form: the reduction is the first ROTA algorithm applied to Display Data.

[0050] Other reduction operations are possible as well. FIG 14 shows image 1400 which illustrates sum reduction 1402 and mean reduction 1404. In one embodiment, the mean (average) reduction is determined using the formula ' ( , y) = G-i Thisreduces the influence of thickness in the boundary region. In one embodiment, gamma i correction is applied to the images as follows: S x, y) -» S (%, y)L Image 1402 shows the second algorithm applying the sum. Each A-Scan line is reduced into a single pixel by applying a summation as a reduction operation. Image 1404 shows the result when the mean (sum over number of elements) is used as the reduction operation for each A- Scan Line.

[0051] Some examples of other statistical reductions include mean, maximum, minimum, standard deviation, and coefficient of variance. FIG. 15 shows a variety of projections (e g., reductions). Image 1502 of FIG. 15 shows a sum reduction with y=1.6. Image 1504 of FIG. 15 shows a min reduction with y=1.6. Image 1506 of FIG. 15 shows a mean reduction with y=1.2. Image 1508 of FIG. 15 shows a standard deviation with y=1.2. Image 1510 of FIG. 15 shows a max reduction with y=1.2. Image 1512 of FIG. 15 shows a coefficient of variance with y=1.0.

[0052] In one embodiment, 3D data representing an object (e.g., a patient’s eye) is received and is converted into 2D data representing a projection of the object based on at least one of a linear scale reflected intensity value or a derived quantity. The 2D data is then used to train or tune an image processing algorithm. In one embodiment, an image processing algorithm is trained (e.g., the image processing algorithm is initially trained) and tuned (e.g., modification of the image processing algorithm after initial training). In one embodiment, the 3D data comprises optical coherence tomography data and the derived quantity is an attenuation coefficient. In one embodiment, the linear scale reflected intensity value is generated based on spectrometer data (e.g., SD OCT: spectrometer data, SS OCT: digitizer, or interferogram data can be commonly used).Generating reference data based on optical coherence tomography data

[0053] FIG. 16 shows flow diagrams 1600A and 1600B through 1600X for generating reference data based on received optical coherence tomography data (e.g., optical coherence tomography data of a healthy eye used as a reference eye). 3D optical coherence tomography data 1602 and segmentation data 1604 are used to generate 3D slab data 1606. In one embodiment, the 3D optical coherence tomography data is received and is segmented into a plurality of retinal layers. 3D slab data 1606 is then converted by 2D transformation 1608 to generate 2D data 1610. In one embodiment, the 2D data is generated as a transformation from a reflected amplitude (e.g., reflected amplitude of reflectance data) of the optical coherence tomography data (e.g., by applying a transformation operation). Other data 1612 (e.g., patient demographic, age, gender, existing diagnoses, ethnicity, blood type, BMI, HbAlC, etc.) is used statistical calculation. 2D data 1610 can be used for image registration 1614 in order to align 2D data 1610 and other data 1612, with each other as appropriate (i.e., data that pertains to image registration). After image registration 1614, statistical calculation 1616 is performed to produce reference data 1618. In one embodiment, the statistical calculation calculates an epidemiological distribution of pixel data in the 2D data. In one embodiment, the statistical analysis comprises a multi-variable analysis that is based onone or more of a pixel value of the reflectance data, a patient age, a disc size, a disc morphology, a fovea-to-disc distance, or vessel patterns

[0054] Method 1600B is the same method as method 1600A but can use different optical coherence tomography data. Additional iterations of “X” number of methods used with different optical coherence tomography data in order to generate data that can be used to generate reference data. The different optical coherence tomography data can be from different people, different eyes, and / or one or more of a plurality of retinal layers / slabs of data from a particular eye.

[0055] FIG. 17 shows how 3D optical coherence tomography data 1602 and segmentation data 1604 are used to generate reference data for various slabs / layers of a patient’s eye. RNFL slab 1702 is generated based on tomography data 1602 and segmentation data 1604. RNFL slab 1702 is transformed to RNFL 2D data 1704 and image registration 1706 is performed as necessary before statistical calculation 1708 is used to generate RNFL 2D reference data 1710. Similarly, GCL slab 1712 is generated based on tomography data 1602 and segmentation data 1604 and is transformed to GCL 2D data 1714 and image registration 1716 is performed as necessary to align the images of the patient’s eye with reference data before statistical calculation 1718 is used to generate GCL 2D reference data 1720. IPL slab 1722 is generated based on tomography data 1602 and segmentation data 1604 and is transformed to IPL 2D data 1724 and image registration 1726 is performed as necessary to align the images of the patient’s eye with reference data before statistical calculation 1728 is used to generate IPL 2D reference data 1722. ILM-RPE slab 1732 is generated based on tomography data 1602 and segmentation data 1604 and is transformed to ILM-RPE 2D data 1734 and image registration 1736 is performed as necessary to align the images of the patient’s eye with reference data before statistical calculation 1738 is used to generate ILM-RPE 2D reference data 1732.

[0056] 2D reference data 1710, 1720, 1730, and 1740, along with other data that was used to generate the 2D reference data can be stored in a data base. The data in the database can be accessed for analysis and / or review and can be used to diagnose various conditions and / or diseases.Method and apparatus for calculating a health risk score

[0057] Ophthalmological images can be used to diagnose or track a patient’s condition and / or disease. Figure 18 shows flow diagram 1800 for calculating a health risk score based on optical coherence tomography data. OCT data 1802 comprises 3D optical coherence tomography data of the eye obtained by scanning the eye of a subject using a device, such as an OCT device. The OCT device may be any suitable device that can acquire 3D images of the retina, such as a spectral-domain OCT (SD-OCT), a swept- source OCT (SS-OCT), or a full-field OCT (FF-OCT). The 3D optical coherence tomography data of eye may include information about the intensity, the phase, the polarization, or the speckle of the light reflected from the retinal layers.

[0058] 3D optical coherence tomography data 1804 and segmentation data 1806 from received OCT data 1802 are used to generate 3D slab data 1808 by segmenting the 3D data into a plurality of retinal layers. The 3D data may be segmented in each of the plurality of retinal layers using any suitable segmentation algorithm, such as a graphbased algorithm, a region-based algorithm, a level-set algorithm, a deep learning algorithm, or a combination thereof. The segmentation algorithm may identify and separate the boundaries of different retinal layers, such as the nerve fiber layer (NFL), the ganglion cell layer (GCL), the inner plexiform layer (IPL), the inner nuclear layer (INL), the outer plexiform layer (OPL), the outer nuclear layer (ONL), the photoreceptor layer (PRL), the retinal pigment epithelium (RPE), or the choroid layer (CL). The segmentation algorithm may also segment the 3D data in the macular region, the optic disc region, the fovea region, the vascular region, or any other region of interest.

[0059] 3D slab data 1808 is then transformed at 2D transformation 1810 to generate 2D data 1820. The 3D segmented data may be converted into a 2D image using a transformation of the data between specified layer boundaries. The specified layer boundaries may be any two boundaries that define a retinal layer or a sub-layer, such as the NFL boundary, the GCL boundary, the IPL boundary, the INL boundary, the OPL boundary, the ONL boundary, the PRL boundary, the RPE boundary, or the CL boundary. Specified layer boundaries can also span multiple layers. For example, an upper boundary can be defined as the NFL boundary and the lower boundary can bedefined as the OPL boundary. A specified boundary can also be defined as an offset(s) from an existing boundary. For example, layer can be defined as being bound by the NFL boundary less lOum and the IPL boundary plus 15.2um. The transformation may be any suitable transformation that can map the 3D data between the specified layer boundaries onto a 2D plane, such as a projection, a flattening, a warping, a stretching, a shrinking, a folding, a rolling, or a combination thereof. The transformation may preserve the information of the 3D data, such as the intensity, the phase, the polarization, or the speckle, in the 2D image. The transformation may also enhance the contrast, the resolution, the sharpness, or the visibility of the 2D image.

[0060] Comparison 1822 is performed by comparing 2D data 1820 with reference data 1824. In one embodiment, reference data 1824 comprises data corresponding to the 2D data 1820 for comparison. For example, reference data 1824 can be images representing a disease-free layer of an eye that can be compared with layers generated from optical coherence tomography data of a patient’s eye in order to identify abnormalities. 2D data 1820 and reference data 1824 may be compared using any suitable comparison algorithm, such as a pixel-wise algorithm, a feature-based algorithm, a histogram-based algorithm, a template-matching algorithm, a similarity-measuring algorithm, a difference-detecting algorithm, or a combination thereof. 2D data and reference data can be compared to identify abnormalities. The comparison algorithm may calculate a comparison result that indicates the degree of similarity, dissimilarity, agreement, disagreement, matching, mismatching, or deviation between the 2D data and the reference data. The comparison result may be a numerical value, a vector, a matrix, a tensor, a score, a ratio, a percentage, a probability, a distribution, a map, a graph, or any other representation that can quantify or visualize the comparison.

[0061] A result of comparison 1822 along with color fundus data 1828 and historical data 1830 (e.g., patient history data) are then input to trained machine learning model 1826 which generates one or more risk scores (described in detail below) based on the result (e g., a health risk score generated based on identified abnormalities). The trained machine learning model may be any suitable machine learning model that can perform a regression, a classification, a clustering, a dimensionality reduction, a feature extraction,a feature selection, a feature engineering, a feature learning, or a combination thereof The trained machine learning model may be a supervised machine learning model, an unsupervised machine learning model, a semi-supervised machine learning model, a reinforcement machine learning model, or a combination thereof. The trained machine learning model may be a linear model, a logistic model, a support vector machine, a decision tree, a random forest, a k-nearest neighbor, a k-means, a principal component analysis, a neural network, a convolutional neural network, a recurrent neural network, a long short-term memory network, a generative adversarial network, a transformer, a BERT, a GPT, or a combination thereof. The trained machine learning model may be trained using a training data set that comprises a plurality of pairs of comparison results and health risk scores. The training data set may be obtained from a database, a repository, a library, or any other source that stores comparison results and health risk scores of subjects. The training data set may be pre-processed, cleaned, augmented, balanced, or shuffled before the training. The trained machine learning model may be validated, tested, evaluated, or optimized using a validation data set, a test data set, a cross-validation data set, or any other data set that can measure the performance, the accuracy, the precision, the recall, the sensitivity, the specificity, the Fl -score, the ROC curve, the AUC, the MSE, the MAE, the RMSE, the R-squared, or any other metric of the trained machine learning model.

[0062] Trained machine learning model 1826 generates a risk score for neurological disease 1832, a risk score for systemic disease 1834, a risk score for ophthalmic disease 1836, and a risk score for other (diseases) 1838. These risk scores can then be displayed to a user via display 1840.

[0063] In one embodiment, the methods, systems, and techniques described herein are performed on a computer. A high-level block diagram of such a computer is illustrated in FIG.19. Computer 1902 contains a processor 1904 which controls the overall operation of the computer 1902 by executing computer program instructions which define such operation. The computer program instructions may be stored in a storage device 1912, or other computer readable medium (e.g., magnetic disk, CD ROM, etc.), and loaded into memory 1910 when execution of the computer program instructions isdesired. Thus, the methods, systems, and techniques described herein, can be defined by the computer program instructions stored in the memory 1910 and / or storage 1912 and controlled by the processor 1904 executing the computer program instructions. For example, the computer program instructions can be implemented as computer executable code programmed by one skilled in the art to implement an algorithm defined by the methods and techniques described herein. Accordingly, by executing the computer program instructions, the processor 1904 executes an algorithm defined by the methods and techniques described herein. The computer 1902 also includes one or more network interfaces 1906 for communicating with other devices via a network. The computer 1902 also includes input / output devices 1908 that enable user interaction with the computer 1902 (e.g., display, keyboard, mouse, speakers, buttons, etc.) In one embodiment, computer 1902 can include additional components such as processors. For example, computer 1902 can include a general purpose graphics processing unit (GPGPU) and a significant portion of the calculations described herein can be performed by the GPGPU. One skilled in the art will recognize that an implementation of an actual computer could contain other components as well, and that FIG. 19 is a high-level representation of some of the components of such a computer for illustrative purposes.

[0064] The foregoing Detailed Description is to be understood as being in every respect illustrative and exemplary, but not restrictive, and the scope of the inventive concept disclosed herein is not to be determined from the Detailed Description, but rather from the claims as interpreted according to the full breadth permitted by the patent laws. It is to be understood that the embodiments shown and described herein are only illustrative of the principles of the inventive concept and that various modifications may be implemented by those skilled in the art without departing from the scope and spirit of the inventive concept. Those skilled in the art could implement various other feature combinations without departing from the scope and spirit of the inventive concept.

Claims

CLAIMS:

1. A method for assessing the health of a patient’s eye, the method comprising: receiving three-dimensional optical coherence tomography data of a reference eye; segmenting the three-dimensional optical coherence tomography data into a plurality of retinal layers; and converting each of the plurality of retinal layers into two-dimensional reference data by applying a transformation operation.

2. The method of claim 1, further comprising: identifying abnormalities in the patient’s eye by comparing the two-dimensional reference data to two-dimensional patient eye data.

3. The method of claim 1, further comprising: calculating an epidemiological distribution of pixel data in the two-dimensional reference data.

4. The method of claim 1, wherein the two-dimensional reference data is generated as a transformation from a reflected amplitude of the three-dimensional optical coherence tomography data.

5. The method of claim 4, wherein the transformation is one of ROTA, sum, average, mean, standard deviation, or coefficient of variance.

6. The method of claim 3, wherein the calculating an epidemiological distribution of pixel data in the two-dimensional reference data comprises a multi-variable analysis.

7. The method of claim 6, wherein the multi-variable analysis is based on one or more of a patient gender, a pixel value of reflectance data, a patient age, a disc size, a disc morphology, a fovea-to-disc distance, or vessel patterns.

8. The method of claim 2, further comprising: calculating a health risk score by a machine learning model based on the identified abnormalities.

9. The method of claim 8, wherein the calculating the health risk score is further based on color fundus data.

10. The method of claim 8, wherein the health risk score is related to at least one of neurological disease, systemic disease, or ophthalmic disease.

11. The method of claim 1, wherein each of the plurality of retinal layers are identified by a patient gender and by one of a macular region, an optic disc region, a fovea region, or a vascular region.

12. An apparatus for assessing the health of a patient’s eye, the apparatus comprising: a processor; and a memory to store computer program instructions, the computer program instructions, which, when executed on the processor cause the processor to perform operations comprising: receiving three-dimensional optical coherence tomography data of a reference eye; segmenting the three-dimensional optical coherence tomography data into a plurality of retinal layers; and converting each of the plurality of retinal layers into two-dimensional reference data by applying a transformation operation.

13. The apparatus of claim 12, the operations further comprising: identifying abnormalities in the patient’s eye by comparing the two-dimensional reference data to two-dimensional patient eye data.

14. The apparatus of claim 12, the operations further comprising: calculating an epidemiological distribution of pixel data in the two-dimensional reference data.

15. The apparatus of claim 12, wherein the two-dimensional reference data is generated as a transformation from a reflected amplitude of the three-dimensional optical coherence tomography data.

16. The apparatus of claim 13, the operations further comprising: calculating a health risk score by a machine learning model based on the identified abnormalities.

17. The apparatus of claim 16, wherein the calculating the health risk score is further based on color fundus data.

18. The apparatus of claim 16, wherein the health risk score is related to at least one of neurological disease, systemic disease, or ophthalmic disease.

19. A computer readable medium storing computer program instructions for assessing the health of a patient’s eye, which, when executed on a processor, cause the processor to perform operations comprising: receiving three-dimensional optical coherence tomography data of a reference eye; segmenting the three-dimensional optical coherence tomography data into a plurality of retinal layers; andconverting each of the plurality of retinal layers into two-dimensional reference data by applying a transformation operation.

20. The computer readable medium of claim 19, the operations further comprising: identifying abnormalities in the patient’s eye by comparing the two-dimensional reference data to two-dimensional patient eye data.

21. The computer readable medium of claim 19, the operations further comprising: calculating an epidemiological distribution of pixel data in the two-dimensional reference data.

22. The computer readable medium of claim 19, wherein the two-dimensional reference data is generated as a transformation from a reflected amplitude of the three- dimensional optical coherence tomography data.

23. The computer readable medium of claim 20, the operations further comprising: calculating a health risk score by a machine learning model based on the identified abnormalities.

24. The computer readable medium of claim 23, wherein the calculating the health risk score is further based on color fundus data.

25. The computer readable medium of claim 23, wherein the health risk score is related to at least one of neurological disease, systemic disease, or ophthalmic disease.

Images