WO2022037517A1

WO2022037517A1 - Methods and system for automated cloud-based quantitative assessment of retinal microvasculature using optical coherence tomography angiography images

Info

Publication number: WO2022037517A1
Application number: PCT/CN2021/112735
Authority: WO
Inventors: Kin CHIU; Bingwen LU; Yiming Li; Pak Chung SHAM
Original assignee: The University Of Hong Kong
Priority date: 2020-08-17
Filing date: 2021-08-16
Publication date: 2022-02-24
Also published as: CN116157829A

Abstract

Systems and methods for the automated quantitative assessment of a retinal microvasculature and quantification of microvascular characteristics are provided for the diagnosis and treatment of retinal and extra-retinal diseases.

Description

METHODS AND SYSTEM FOR AUTOMATED CLOUD-BASED QUANTITATIVE ASSESSMENT OF RETINAL MICROVASCULATURE USING OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY IMAGES

TECHNICAL FIELD

The present invention is directed to methods and system for automated cloud-based quantitative assessment of retinal microvasculature using optical coherence tomography angiography images.

BACKGROUND

Retinal diseases in humans can be manifestations of different physiological or pathological conditions including diabetes that causes diabetic retinopathy (DR) , retinal vein occlusion (RVO) , age-related macular degeneration (AMD) , and others. Of late, improved longevity and “stationary” , stress-filled lifestyles have resulted in a rapid increase in the number of patients suffering from these vision threatening conditions. There exists an urgent need for a large-scale improvement in the way in which these diseases are screened, diagnosed and treated.

Retinal imaging is the only way to directly inspect blood vessels and their pathological changes throughout the whole body. Retinal imaging is capable of not only reflecting retinal vascular disorders but also indicates risks of systemic diseases including as cardiovascular disorders and hypertension. A majority of retinal diseases lead to abnormality in microvasculature beds in the eye, including DR, RVO, and AMD. Furthermore, retinal characteristics provide important information on the risk of developing vascular disorders. Systemic diseases such as cardiovascular diseases, hypertension, and atherosclerosis manifest themselves in retinal microvascular changes. Therefore, the measurement of retinal microvasculature plays a pivotal role in the management of vascular disorders.

Optical coherence tomography angiography (OCTA) has been recently introduced for the imaging of microvascular networks in the human eye. Compared with conventional dye-based angiography, OCTA offers a safer, faster, and more cost-effective non-invasive way to diagnose and monitor retinal vascular abnormalities. OCTA has the ability to show the microvasculature of the superficial, deep and avascular retinal maps as well as choriocapillaris and choroid maps.

Typically, patients visit an ophthalmologist or other eye care professionals on a regular basis to go through OCTA measurements. The professional may inspect the images and intellectually compare these to other retinas they have observed in the past. This experience-based examination, while effective at detecting gross abnormalities and large changes, may not be effective in observing small changes in the patient’s retina. The evaluation is also subjective and its accuracy leans heavily on the prior experience of the clinician, making the diagnosis prone to human error. It is not uncommon for a less experienced ophthalmologist to consult external experts on the interpretation of lab results and medical images to improve their diagnosis accuracy. On the other hand, a majority of patients do not have access to well-known ophthalmology clinicians and cannot attain enough information directly from their OCTA reports due to limited descriptive capabilities of the OCTA machine’s built-in software. When a patient is hence concerned about their eye related health, the patient may have to seek further advice from an external specialist, which can be time-consuming and expensive. Therefore, current retinal analysis systems are subjective, time-consuming, prone to error and require significant manual input from the medical personnel. Also, comprehensive quantitative OCTA-image-based biomarkers to reliably monitor vascular diseases are missing.

Advances in computer-aided image processing and analysis technologies are essential for making imaging-based disease diagnosis scalable, cost-effective, and reproducible. Such advances would directly result in effective triage of patients, leading to early diagnosis, timely treatment and improved quality of life.

SUMMARY

Provided are systems and methods for automated quantitative assessment of a retinal microvasculature and generation of retinal digital maps for the diagnosis and treatment of retinal and extra-retinal diseases. The systems and methods of the invention provide computational retinal microvascular biomarkers (CRMBs) that can be used in the quantification of retinal microvascular disease states. Advantageously, the systems and methods of the invention can be integrated in cloud computing platforms for cloud-based analytics that provide broad accessibility to patients and health care providers and enable the use of machine learning for the identification of novel CRMBs.

BRIEF SUMMARY OF THE DRAWINGS

Figure 1 shows a schematic diagram of an exemplary architecture of the cloud platform for a computerized automated optical coherence tomography angiography (OCTA) image analysis system.

Figure 2 shows a schematic diagram of the flow of process for extracting computational retinal microvascular biomarkers (CRMBs) from an OCTA image in accordance with embodiments of the present technology.

Figure 3A shows representative OCTA raw original images over both deep and superficial capillary plexuses. Figure 3B shows representative OCTA scanned images over both deep and superficial capillary plexuses.

Figure 4 shows an illustration of the algorithm for extracting and generating a digital vasculature map (DVM) from a calibrated OCTA image.

Figure 5A shows an illustration of the process of applying the early treatment of diabetic retinopathy (ETDRS) grid to a DVM. Figure 5B shows sections defined by the ETDRS grid in the left and the right eye.

Image A of Figure 6 shows a digital vascular map (DVM) generated using the OCTA images. Image B of Figure 6 shows a thresholded image of the DVM image. Image C of Figure 6 shows a thresholded parafoveal image of the DVM image. Image D of Figure 6 shows a ridge filtered image of the DVM image. Image E of Figure 6 shows a thresholded ridge filtered image of DVM image. Image F of Figure 6 shows the image E of Figure 6 with small vessel objects removed. Image G of Figure 6 shows the image F of Figure 6 with retinal vessels delineated. Image H of Figure 6 shows the image F of Figure 6 with skeletonized vessels. Image I of Figure 6 shows the image H of Figure 6 with separated major vessel segments.

Figure 7 shows an illustration of the process for calculating a vessel dispersion index of a DVM.

Figure 8 shows an illustration of the algorithm for detecting endpoints and branch points in a skeletonized DVM.

Figure 9A shows an image of a DVM with the calculated vessel dispersion indices of individual vessels labeled. Figure 9B shows an image of a DVM with the calculated vessel diameter indices of delineated individual vessels labeled. Figure 9C shows an image of a DVM with the calculated vessel tortuosity indices of separated individual vessel segments labeled.

Image A of Figure 10 show a digital vascular map (DVM) of a retina. Image B of Figure 10 shows a masked and denoised DVM. Image C of Figure 10 shows a thresholded and closed DVM. Image D of Figure 10 shows a bridged DVM. Image E of Figure 10 shows a DVM with the foveal avascular zone polygon endpoints. Image F of Figure 10 shows a DVM with the foveal avascular zone demarcated.

Figure 11 shows a visualization of the distribution and pairwise relationships among different CRMBs across samples of healthy subjects and retinal vein occlusion patients.

DETAILED DESCRIPTION

Provided are systems and methods for the automated assessment of retinal microvascular architecture, quantification of retinal microvasculature changes related to retinal or extra-retinal diseases, treatment of the diseases, and quantitative assessment of the treatment success based on the reversal of disease-related changes in the retinal microvasculature. Using the systems and methods of the invention, further provided are methods for establishing comprehensive computational retinal microvasculature biomarkers (CRMBs) through a knowledge-driven computerized automatic analytical system based on fractal analysis using OCTA images. The systems and methods of the invention employ cloud-computing networks with distributed input devices and cloud-based analytics platforms for the automatic analyses of OCTA images, quantification of eye disorders based on CRMBs established using the systems and methods of the invention and treatment of the respective eye disorders. Cloud computing refers to the access of computing resources through the Internet for purposes of data storage, aggregation, synthesis, and retrieval, together with the capacity to act on the data with computational algorithms and software packages.

In preferred embodiments, the invention provides a computerized automated OCTA image analysis system for the development and quantification of computational retinal microvasculature biomarkers (CRMBs) based on OCTA image databases and assessment and quantification of retinal characteristics in a comprehensive and easily accessible manner. Using the systems of the invention, clinicians and patients can obtain and assess data reports within minutes, thus, enabling comprehensive diagnosis and treatment in a clinical setting or an out-patient office visit. The fast and comprehensive analysis also enables seamless re-imaging in cases where conclusive results could not be obtained using the initial OCTA reports.

Advantageously, the system and generated quantitative retinal microvasculature biomarkers using the systems and methods of the invention are also available for effective downstream research efforts.

In some embodiments, the cloud database of the system of the invention stores images originating from thousands of patients, which images have been uploaded either by the patients themselves or collected as part of a clinical study for disease or drug discovery research or resulted from routine clinical workflow in which the analysis was carried out in batch mode for a group of patients.

In preferred embodiments, the data obtained using the systems of the invention are provided to an e-cloud deep learning system that is configured to assess and identify early biomarkers of retinal disease including, but not limited to, retinal vascular disease through machine learning techniques.

Using the computational microvascular biomarkers generated using the methods and systems of the invention, treatment plans for medical conditions associated with and/or underlying the presence of retinal microvascular biomarkers can be devised and patients can be treated and treatment efficiency be assessed using the CRMBs of the invention.

In some embodiments, the CRMBs quantified are used to quantify the extent of an extra-retinal disease condition. In some embodiments, the CRMBs quantified are used to quantify the extent of a retinal disease condition. In some embodiments, the CRMBs quantified are used to quantify both an extra-retinal and a retinal disease condition.

In some embodiments, the systems of the instant invention are integrated into the interface of an OCTA machine and the user can view the results of real-time disease biomarker quantification after OCTA images capture in an eye clinic.

In some embodiments, the systems of the invention provide machine learning techniques to develop quantification parameters for a variety of extra-retinal and retinal disease conditions. In specific embodiments, the machine learning techniques include, but are not limited, to deep learning and gradient boost decision trees. Advantageously, the systems and methods of the invention provided are used to validate and assess novel objective microvasculature biomarkers for eye diseases, thereby facilitating the diagnosis, monitoring and treatment of the respective eye diseases.

In preferred embodiments, the systems and methods quantify retinal microvascular parameters that include but are not limited to, fractal dimension index, foveal avascular zone area, and flow index. In further preferred embodiments, the systems and methods quantify retinal microvascular parameters including retinal vascular geometric characteristics that include, but are not limited to, vessel dispersion, vessel diameter, and vessel tortuosity.

In further embodiments, machine learning techniques including, but not limited to, e-cloud deep learning are provided to identify and quantify novel CRMBs.

In specific embodiments, the systems of the invention acquire OCTA image data. For example, in some embodiments of cloud-based operation of the systems of the invention, patient OCTA images including, but not limited to, two consecutive 3×3 mm, en face OCTA images of a superficial and a deep capillary plexus are generated using, e.g., an OPTOVUE, Cirrus 5000 and/or SPECTRALIS OCT2 OCTA machine at medical center. In other embodiments, scanned images from outputs of OCTA machines are generated. Advantageously, data input can be performed by ophthalmologist users and patients both of which can access the cloud platform and scan OCTA reports from hospital visits to upload into the systems of the invention. The OCTA images are uploaded to the cloud through an Internet application programming interface (API) . In some embodiments, the OCTA images are uploaded as collected in the course of a research study and/or a clinical study for disease and/or drug development. In some embodiments, the OCTA images are uploaded in the course of routine clinical workflow in batch mode or by patients themselves.

In some embodiments, the uploaded OCTA images are anonymized by removing certain types of metadata and stored in the cloud with server-side encryption. Advantageously, the cloud infrastructure accommodates management and analysis of large-scale imaging data using the systems of the invention.

In some embodiments, the systems of the invention are configured to calibrate the raw OCTA images in order to input the images in a predetermined orientation. For example, when OCTA raw images are scanned from a printed output of an OCTA machine and the images are in portrait view the systems of the invention rotated the images by 90 degrees.

In some embodiments, scanned images are converted from RGB (R, G, B) to Lab color space (L, a, b) and equally divided into a left half and a right half.

In specific embodiments, the pixels in the respective image halves are analyzed and images that comprise more pixels with

the left half of the OCTA image are rotated clockwise.

In other embodiments, images that comprise less pixels with

the left half of the OCTA image are rotated counter-clockwise. Following the respective rotations the scanned images are in landscape view and in the predetermined orientation.

In preferred embodiments, the images following rotation to the predetermined orientation are further rotated for a small amount of degrees such that the reference lines are strictly horizontal and vertical, respectively.

In preferred embodiments, the images following rotation to the predetermined orientation are further rotated for the number of degrees needed (ranging from -90 to 90 degrees) such that the reference lines are strictly horizontal and vertical, respectively.

In some embodiments, the scanned images are converted to grayscale, Gaussian-blurred to reduce noise, thresholded and edged through a Canny edge detector.

In preferred embodiments, the rotation degrees are determined by detecting the orientation of reference lines in the edged image utilizing Hough line transform.

In further embodiments, a digital vascular map (DVM) is generated from the OCTA images following the above described calibration process. To this end, the OCTA images are first converted to grayscale and thresholded using Otsu’s binarization. All the vertical and horizontal line segments in the resulting image are identified using Hough line transform and the detected line segments are combined into a new image. The contours in the new image are identified and sorted; the contours correspond to rectangles in the original OCTA image.

In preferred embodiments, the top left rectangle with a width/height larger than 0.27 x the width of the original OCTA image is extracted. In embodiments where the top left rectangle does not have an equal length and width, the rectangle is cropped to a square with cropping determinants determined by performing Harris corner detection on the padded, Gaussian-blurred and thresholded version of the image. The thus generated square of the image is the extracted DVM from the calibrated OCTA image.

In some embodiments, the systems and methods of the invention apply an early treatment of diabetic retinopathy (ETDRS) grid to the DVM generated as described above. In specific embodiments, the systems and methods convert the DVM to Lab color space (L, a, b) and replace all pixels with

with white ones in order to keep only the colored components while removing most of the vessel signals. The resulting image is then converted to grayscale and skeletonized, and the reference circles in the ETDRS grid are identified through Hough circle transform. In specific embodiments, the inner reference circle C _inner has a center (x _inner, y _inner) and a radius of r _inner, whereas the outer reference circle C _outer has a centre (x _outer, y _outer) = (x _inner, y _inner) and a radius of r _outer. In preferred embodiments, the area within C _inner roughly corresponds to the fovea region, whereas the parafovea area refers to the ring region between C _inner and C _outer.

In specific embodiments, the parafovea is subsequently divided into a superior, an inferior, a nasal and a temporal section. In some embodiments, the systems and methods calculate the section boundaries for the right (OD) or left (OS) eye, respectively, using the recorded parameters of the reference circles in the ETDRS grid. In some embodiments, where the coordinates of the top left corner of a DVM are (0, 0) , the end points of the line segment dividing the superior and nasal sections of the right eye are:

and

In some embodiments, the system and methods of the invention calculate a fractal dimension index (FD) to quantify the degree of complexity of the retinal microvascular structure in both the superficial capillary plexus (SCP) and the deep capillary plexus (DCP) . The generated DVM is first binarized through adaptive thresholding, and the FD is calculated by applying the box-counting algorithm to the resulting map (Lemmens et al., 2020) .

In some embodiments, the systems and methods automatically delineate retinal vessels in the SCP. In preferred embodiments, the system and methods binarize the original DVM through adaptive global thresholding resulting in a thresholded DVM image. In preferred embodiments, only the fovea and parafovea regions are retained according to an ETDRS grid applied as described above. In some embodiments, the systems and method of the invention use a Sato’s filter to detect continuous ridges, which correspond to the vessels, in the resulting DVM. The ridge-filtered DVM image is subsequently binarized through adaptive thresholding based on the median of the parafoveal pixel intensity. In further embodiments, the systems and methods remove from the binarized DVM image small objects, which correspond to insignificant isolated vessel branches or noise. In further embodiments, the systems and methods identify contours in the resulting DVM with each contour corresponding to one delineated retinal vessel.

In some embodiments, the capillary perfusion density index (PDC) is calculated for both DCPs and SCPs based on the generated DVM and the ETDRS grid accompanying it. The map is first converted to grayscale with pixel intensity values ranging from 0 (black) to 255 (white) . Next, PDC is calculated as the average intensity of all the non-delineated-vessel pixels in the considered section of the map. Since brighter pixels (with larger intensity values) often correspond to vascular structures in OCTA images, when a map has a greater calculated PDC, it has denser capillaries.

In some embodiments, the large vessels perfusion density index (PDL) is calculated based on the generated DVM in the SCP and its ETDRS grid. For a considered section of the map, PDL is calculated as the proportion of pixels that belong to delineated vessels, or large vessels, in it. When a map has a greater calculated PDL, it has denser large vessels.

In preferred embodiments, the systems and methods calculate a vessel dispersion index (VDisp) . Because in a healthy eye the parafoveal vessels are more centripetal towards the center of the fovea, the VDisp refers to the degree of centripetalism of the parafoveal vessels. The larger the VDisp the less centripetal the vessels are on average.

In some embodiments, the systems and methods, following removal of small objects, pad and crop a thresholded, ridge filtered DVM image such that the parafovea region is centered. In further embodiments, for each delineated retinal vessel V _i (i=1, …, N, where N is the total number of delineated vessels) , the systems and methods calculate an ellipse E _i with a same second moment as the region V _i covers. In some embodiments, if V _i is perfectly centripetal, it is identically orientated as the line segment connecting the centroid of E _i and the center of the fovea.

The VDisp according to methods of the instant invention is defined as the mean of all angles between the major axis of E _i and the line segment connecting the centroid of E _i and the center of C _inner, which corresponds to the fovea identified when applying the ETDRS grid to the extracted and calibrated DVM image. In some embodiments, the systems and methods of the invention calculate a multitude of vessel dispersity indices for individual vessels of a DVM image. In some embodiments, the systems and methods calculate an overall VDisp of the DVM image, which overall VDisp is the average of all individual vessel dispersity indices.

In specific embodiments, the systems and methods separate major and minor retinal vessel segments. In specific embodiments, the systems and methods skeletonize a vasculature map with delineated retinal vessels. In further embodiments, the systems and methods detect endpoints and branch points in the skeletonized DVM image through an algorithm that comprises the steps of: traversing every positive pixel in a DVM image and counting the number of positive pixels (n) in the pixel’s eight neighborhood sections; calculating the number of pixels that contain a single positive pixel in their eight neighborhood sections and applying an endpoint tag to said pixels; calculating the number of pixels that contain three or more positive pixels in their eight neighborhood sections and applying a branch point tag to said pixels; calculating the pixel comprising a branch point tag and which branch point tagged pixels are non-adjacent and applying a true branch point tag to said pixels.

In some embodiments, the systems and methods mask the tagged branch points from the DVM image and calculate all contours in the masked images, where each contour corresponds to one vessel segment. In further embodiments, the systems and methods denote the total number of branch points as N _bp. The total number of vessel segments N _seg is then equal to 1 + 2N _bp. In further embodiments, the systems and methods denote the geodesic length of a segment as L _seg and the geodesic length of the major vessel it belongs to as L _ves.

If

the segment is marked as a minor vessel segment. Otherwise, it is marked as a major vessel segment.

In some embodiments, the systems and methods of the invention calculate the vessel diameter index (VDiam) . The VDiam is defined to be the mean of the ratios between the pixel area and the geodesic length of all the delineated retinal vessels. Denoting S as the side length of the DVM, N as the number of delineated vessels and V _i as the i-th delineated vessel in the image, the formula for VDiam is:

In some embodiments, the overall VDiam of the DVM image is calculated and is the average of all the labelled individual VDiam’s in the image divided by S.

In some embodiments, the systems and methods of the invention calculate the vessel tortuosity index (VT) . The VT is quantified as the mean of the ratios between the geodesic length and Euclidean length of the separated vessel segments of a DVM image after the calculation of endpoints and branch points. Denoting M ₁ as the number of major vessel segments, M ₂ as the number of minor vessel segments, S _i ^major as the i-th major vessel segment and S _i ^minor as the i-th minor vessel segment, the formula for VT is:

Where α=0.8 is a weighting factor. In this VT calculation, because minor vessel segments often correspond to secondary vessels less emphasis is laid on the minor vessel segments. If vessels in a DVM image are more curved and twisty on average, the map will have a larger VT. In some embodiments, the individual VTs are displayed in a DVM image. In some embodiments, an overall VT of the DVM image is displayed and is a weighted average of all the labelled individual VT’s in the DVM.

In some embodiments, the system and methods demarcate the FAZ in the DVM image, which is defined to be the capillary-free area within the innermost ring of the parafoveal capillary plexus. In specific embodiments, the reference circles and lines representing the ETDRS grid in the DVM are masked with a median filtered version of their corresponding areas in the map. In some embodiments, the systems and methods apply median filtering to suppress noise. In preferred embodiments, the masked and denoised image is adaptively thresholded and closed (dilation followed by erosion) small objects are removal. In further embodiments, the systems and methods apply bridging, that is, setting 0-valued pixels to 1 if they have two nonzero neighbors that are not connected and thinning to the resulting map. In yet further embodiments, the systems and methods remove small isolated objects from the image to improve the robustness of the FAZ identification method.

In some embodiments, the systems and methods calculate a maximum polygon in the foveal avascular, i.e., a polygon that does not cover any of the positive vessel signals of the resulting image. In some embodiments, the indication (0, 0) corresponds to the top left corner of a DVM, the center of the fovea region is (x _inner, y _inner) and the side length of the DVM is s. In some embodiments, for Θ = 0, 3, 6, 9, …, 357, i.e., altogether 120 Θ’s all the pixels within the image on the directed line segment from (x _inner, y _inner) to (x _inner+s·sinθ, y _inner+s·cosθ) considered. In some embodiments, the first pixel with a positive signal, i.e., that is part of a vessel is recorded as one of the endpoints of the FAZ. If the values of all the pixels within the image on the directed line segment are zero, this particular Θ is skipped. If less than three searches returned a positive signal, an error is reported. In specific embodiments, the polygon is defined by connecting all the n=120 endpoints in the order they were found is the approximated FAZ. In other embodiments, when less than all endpoints are used where some of the searches did not find any positive pixels, the polygon is defined by connecting the less than 120 endpoints to demarcate the FAZ.

In some embodiments, the systems and methods of the invention calculate the FAZ area index (FAI) . In some embodiments, the systems and methods demarcate in clockwise order the vertices of the FAZ polygon as (a ₁, b ₁) , (a ₂, b ₂) , …, (a _n, b _n) . In specific embodiments, the systems and methods calculate the area of the polygon using the shoelace formula:

In further embodiments, the systems and methods define the FAI to be the FAZ polygon area divided by the pixel area of the DVM. Advantageously, the FAI is comparable across images of different sizes if the DVM is always focused on a 3×3 region.

In some embodiments, the systems and methods of the invention calculate the FAZ perimeter index (FPI) . In some embodiments, the systems and methods calculate the perimeter of the FAZ polygon using the below formula:

In some embodiments, the FPI is defined to be the FAZ polygon perimeter divided by the side length (in pixels) of the DVM. Advantageously, the FPI is comparable across images of different sizes if the DVM is always focused on a 3×3 region.

In some embodiments, the systems and methods of the invention calculate the FAZ acircularity index (FACI) . ) . In specific embodiments, the FACI quantifies the irregularity of the shape of the demarcated FAZ and is calculated by the following formula:

Advantageously, when FACI is equal to one, the FAZ is a perfect circle. In some embodiments, when the FACI value is greater than one, the FAZ has more irregular shapes.

In preferred embodiments, learning algorithms are employed by the systems of the invention together with training data generated using the cloud-based OCTA analysis system to derive model parameters to formulate CRMBs for the specific disease conditions quantified.

In some embodiments, the invention provides a method for automated quantitative assessment of retinal microvasculature and generation of a retinal digital vascular map comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data; and

generating a retinal digital vascular map (DVM) .

In some embodiments, the pre-processing step of said method further comprises applying an Early Treatment Diabetic Retinopathy Study grid to the DVM; calculating a fractal dimension index; and demarcating retinal blood vessels.

In further embodiments, the pre-processing step of said method further comprises calculating a blood vessel dispersion index; separating major and minor blood vessel segments; calculating a blood vessel diameter index; calculating a blood vessel tortuosity index; demarcating a foveal avascular zone; calculating a foveal avascular zone area index; and calculating a foveal avascular zone acircularity index.

In preferred embodiments, the invention provides a method of extracting computational retinal microvascular biomarkers (CRMBs) from a DVM generated using the methods and systems of the invention.

In some embodiments, the CRMB is a fractal dimension index , a blood vessel dispersion index, a capillary perfusion density index, a large vessels perfusion density index, a blood vessel diameter index, a blood vessel tortuosity index, a foveal avascular zone area index a foveal avascular zone perimeter index and/or a foveal avascular zone acircularity index. In preferred embodiments, the CRMB is a blood vessel dispersion index or a blood vessel tortuosity index.

In specific embodiments, the invention provides a method for automated quantification of computational retinal microvascular biomarkers from a retinal superficial capillary plexus of a subject suspected to suffer from a retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from a retinal disease, the method comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

generating a retinal digital vascular map (DVM) of the superficial capillary plexus;

providing a retinal DVM of a normal subject;

quantifying at least one CRMB in the DVM of the subject and the DVM of the normal subject; and treating the subject for a retinal disease if the quantity of the at least one CRMB in the subject exceeds the quantity of said CRMB in the normal subject.

In further embodiments, the pre-processing step of said method comprises applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and the quantifying at least one CRMB comprises quantifying a fractal dimension index (FD) ; quantifying a capillary perfusion density index; quantifying a large vessels perfusion density index; quantifying a blood vessel dispersion index; quantifying a blood vessel diameter index; quantifying a blood vessel tortuosity index; quantifying a foveal avascular zone area index; quantifying a foveal avascular zone perimeter index and/or quantifying a foveal avascular zone acircularity index.

In some embodiments, the invention provides a method for automated quantification of computational retinal microvascular biomarkers from a retinal deep capillary plexus of a subject suspected to suffer from a retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from a retinal disease, the method comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

generating a retinal digital vascular map (DVM) of the deep capillary plexus;

providing a retinal DVM of a deep capillary plexus of a normal subject;

In some embodiments, the pre-processing step of said method comprises applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and the quantifying at least one CRMB comprises quantifying a fractal dimension index; quantifying a capillary perfusion density index; quantifying a large vessels perfusion density index; quantifying a blood vessel dispersion index; quantifying a blood vessel diameter index; quantifying a blood vessel tortuosity index; quantifying a foveal avascular zone area index; quantifying a foveal avascular zone perimeter index and/or quantifying a foveal avascular zone acircularity index.

In specific embodiments, the retinal disease is selected from diabetic retinopathy, retinal vein occlusion, and age-related macular degeneration.

In some embodiments, the treatment comprises administering an antibody to a vascular growth factor, an aptamer to a vascular growth factor, an anti-angiogenic steroid, an anti-oxidant supplementation, a laser photocoagulation therapy, a transpupillary thermotherapy, and/or performing an eye surgery.

In preferred embodiments, the invention provides a method for automated quantification of computational retinal microvascular biomarkers from a retinal superficial capillary plexus of a subject suspected to suffer from an extra-retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from an extra-retinal disease, the method comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

providing a retinal DVM of a normal subject;

quantifying at least one CRMB in the DVM of the subject and the DVM of the normal subject; and treating the subject for an extra-retinal disease if the quantity of the at least one CRMB in the subject exceeds the quantity of said CRMB in the normal subject.

In specific embodiments, the pre-processing step of said method comprises applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and the quantifying at least one CRMB comprises quantifying a fractal dimension index; quantifying a capillary perfusion density index; quantifying a large vessels perfusion density index; quantifying a blood vessel dispersion index; quantifying a blood vessel diameter index; quantifying a blood vessel tortuosity index; quantifying a foveal avascular zone area index; quantifying a foveal avascular zone perimeter index and/or quantifying a foveal avascular zone acircularity index.

In some embodiments, the extra-retinal disease is cardiovascular disease, systemic arterial hypertension, sickle cell anemia, anemia, pre-eclampsia, arteritis, arterial macroaneurysm, diabetes mellitus, and/or a hypercoagulation disorder.

In some embodiments, the treatment comprises administering an anti-hypertensive drug, a cholesterol-modifying drug, aspirin, a beta receptor blocker, an alpha receptor blocker, an alpha-2 receptor agonist, a central agonist, a peripheral adrenergic inhibitor, a calcium channel blocker, nitroglycerin, an angiotensin-converting enzyme inhibitor, an angiotensin II receptor blocker, an anticoagulant, a blood thinning drug, an anti-platelet agent, a digitalis preparation, a diuretic, a vasodilator an iron supplementation, a steroid, an antidiabetic drug, an endovascular stent surgery and/or a vascular bypass surgery.

In some embodiments, the treatment comprises administering a metformin, a sulfonylurea, a meglitinide, a thiazolidinedione, a DPP-4 inhibitor, a GLP-1 receptor agonist, a SGLT2 inhibitor, an insulin, and/or an iron supplement.

In preferred embodiments, the invention provides a system for automated quantitative assessment of retinal microvasculature and generation of a retinal digital vascular map, comprising:

a storage, configured to store one or more executable instructions;

a processor, configured to execute the one or more executable instructions so as to implement the steps:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data; and

generating a retinal digital vascular map (DVM) .

In some embodiments, the pre-processing comprises: applying an Early Treatment Diabetic Retinopathy Study grid to the DVM; calculating a fractal dimension index ; and demarcating retinal blood vessels.

In some embodiments, the pre-processing further comprises: calculating a blood vessel dispersion index; separating major and minor blood vessel segments; calculating a blood vessel diameter index; calculating a blood vessel tortuosity index; demarcating a foveal avascular zone; calculating a foveal avascular zone area index; and calculating a foveal avascular zone acircularity index.

In preferred embodiments, the invention provides a system for extracting computational retinal microvascular biomarkers (CRMBs) from a DVM, comprising:

a storage, configured to store one or more executable instructions;

a processor, configured to execute the one or more executable instructions so as to extract computational retinal microvascular biomarkers (CRMBs) from the aforementioned DVM.

In some embodiments, the CRMB is selected from blood fractal dimension index, capillary perfusion density index, large vessels perfusion density index, blood vessel dispersion index, blood vessel diameter index, blood vessel tortuosity index, foveal avascular zone area index, foveal avascular zone perimeter index and foveal avascular zone acircularity index.

In preferred embodiments, the invention provides a system for automated quantification of computational retinal microvascular biomarkers from a retinal superficial capillary plexus of a subject suspected to suffer from a retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from a retinal disease, the system comprising:

a storage, configured to store one or more executable instructions;

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

providing a retinal DVM of a normal subject;

quantifying at least one CRMB in the DVM of the subject and the DVM of the normal subject; and

treating the subject for a retinal disease if the quantity of the at least one CRMB in the subject exceeds the quantity of said CRMB in the normal subject.

In some embodiments, the pre-processing comprises: applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and wherein the quantifying at least one CRMB comprises: quantifying a fractal dimension index; quantifying a capillary perfusion density index; quantifying a large vessels perfusion density index; quantifying a blood vessel dispersion index; quantifying a blood vessel diameter index; quantifying a blood vessel tortuosity index; quantifying a foveal avascular zone area index; quantifying a foveal avascular zone perimeter index and/or quantifying a foveal avascular zone acircularity index.

In some embodiments, the retinal disease is selected from diabetic retinopathy, retinal vein occlusion, and age-related macular degeneration and the treatment comprises administering an antibody to a vascular growth factor, an aptamer to a vascular growth factor, an anti- angiogenic steroid, an anti-oxidant supplementation, a laser photocoagulation therapy, a transpupillary thermotherapy, and/or performing an eye surgery.

In preferred embodiments, the invention provides a system for automated quantification of computational retinal microvascular biomarkers from a retinal deep capillary plexus of a subject suspected to suffer from a retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from a retinal disease, the system comprising:

a storage, configured to store one or more executable instructions;

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

generating a retinal digital vascular map (DVM) of the deep capillary plexus;

providing a retinal DVM of a deep capillary plexus of a normal subject;

In some embodiments, the pre-processing comprises applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and the quantifying at least one CRMB comprises: quantifying a fractal dimension index; quantifying a capillary perfusion density index; quantifying a large vessels perfusion density index; quantifying a blood vessel dispersion index; quantifying a blood vessel diameter index; quantifying a blood vessel tortuosity index; quantifying a foveal avascular zone area index; quantifying a foveal avascular zone perimeter index and/or quantifying a foveal avascular zone acircularity index.

In some embodiments, the retinal disease is selected from diabetic retinopathy, retinal vein occlusion, and age-related macular degeneration and the treatment comprises administering an antibody to a vascular growth factor, an aptamer to a vascular growth factor, an anti-angiogenic steroid, an anti-oxidant supplementation, a laser photocoagulation therapy, a transpupillary thermotherapy, and/or performing an eye surgery.

In preferred embodiments, the invention provides a system for automated quantification of computational retinal microvascular biomarkers from a retinal superficial capillary plexus of a subject suspected to suffer from an extra-retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from an extra-retinal disease, the system comprising:

a storage, configured to store one or more executable instructions;

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

providing a retinal DVM of a normal subject;

treating the subject for an extra-retinal disease if the quantity of the at least one CRMB in the subject exceeds the quantity of said CRMB in the normal subject.

In some embodiments, the extra-retinal disease is selected from cardiovascular disease, systemic arterial hypertension, sickle cell anemia, anemia, pre-eclampsia, arteritis, arterial macroaneurysm, diabetes mellitus, and/or a hypercoagulation disorder and the treatment comprises administering an anti-hypertensive drug, a cholesterol-modifying drug, aspirin, a beta receptor blocker, a calcium channel blocker, nitroglycerin, an angiotensin-converting enzyme inhibitor, an angiotensin II receptor blocker, a blood thinning drug; an iron supplementation, a steroid, an antidiabetic drug, an endovascular stent surgery and/or a vascular bypass surgery.

In preferred embodiments, the invention provides a computer-readable storage medium, which contains a computer program, and the computer program can be executed by a processor to implement the steps of the method of the present invention.

MATERIALS AND METHODS

All patents, patent applications, provisional applications, and publications referred to or cited herein are incorporated by reference in their entirety, including all figures and tables, to the extent they are not inconsistent with the explicit teachings of this specification.

Following are examples that illustrate procedures for practicing the invention. These examples should not be construed as limiting. All percentages are by weight and all solvent mixture proportions are by volume unless otherwise noted.

EXAMPLE 1-SYSTEMS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS

An automated analysis system for OCTA image analysis using a cloud platform was designed (FIG. 1) . The system was designed to be accessed by ophthalmologists and researchers at medical centers.

To this end, patient OCTA images comprising two consecutive 3×3 mm OCTA en face images of superficial and deep capillary plexuses were obtained from different OCTA machines including OPTOVUE, Cirrus 5000 and SPECTRALIS OCT2 that were available at the medical center. In addition to ophthalmologists, patients could also access the cloud platform of the system and scan their OCTA reports from hospital visits. The OCTA images were uploaded to the cloud through an Internet application programming interface (API) , were anonymized by removing certain types of metadata and stored in the cloud with server-side encryption. The cloud infrastructure could accommodate, manage and analyze large-scale imaging data. Advantageously, the computerized automated OCTA image analysis system of the invention determined CRMBs based on the images and returned reports to the users within minutes. Thus, during a patient’s clinic visit the clinician could discuss the OCTA results with the patient. Further, the system enabled seamless re-imaging in cases where conclusive results could not be obtained using the initial OCTA reports. Also, users in research could perform effective downstream statistical analysis for the quantitative assessment of biomarkers. Advantageously, the cloud database enabled the storage of images from thousands of patients including images from data uploaded by the patients themselves, collected as part of a clinical study for disease or drug discovery research, or collected during routine clinical workflow where the analysis was carried out in batch mode for a group of patients. The data obtained using the instant systems was used to develop predictive features through machine learning techniques including deep learning and gradient boost decision trees.

EXAMPLE 2-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF FRACTAL DIMENSION

A process that extracted different CRMBs in compliance with embodiments of the present technology was generated (FIG. 2, block [1] ) . The process acquired OCTA image data (FIG. 2 block [2] and FIG. 3A-3B) . The process calibrated each raw OCTA image such that it was in the correct orientation (FIG. 2, block [3] ) . This was often required when the raw image was a scanned version of a printed output of the OCTA machine. For example, when the image was in portrait view it was rotated by 90 degrees. The scanned image was then converted from RGB (R, G, B) to Lab colour space (L, a, b) and equally divided into a left half and a right half. If the left half had more pixels with

he image was rotated clockwise. Otherwise, the image was rotated counter-clockwise. After rotation, all scanned images were in landscape view and in the correct orientation.

The thus oriented images were further rotated for a small amount of degrees in order to bring the reference lines as strictly horizontal/vertical. Subsequently, the scanned image was converted to grayscale, Gaussian-blurred to reduce noise, thresholded and edged through the Canny edge detector. Then, the rotation degrees were determined by detecting the orientation of reference lines in the edged image utilizing Hough line transform.

The process extracted and generated a digital vasculature map (DVM) from the calibrated OCTA image (FIG. 2, block [4] ) . To this end, the OCTA image was first converted to grayscale and thresholded using Otsu’s binarization. All the vertical/horizontal line segments in the resulting image were identified through Hough line transform and the detected line segments were combined into a new image. The contours in the new image corresponded to rectangles in the original OCTA image. The top left rectangle with a width/height larger than 0.27 × the width of the original OCTA image was extracted. If the rectangle did not have an equal length and width, it was cropped to square with cropping coordinates determined by performing Harris corner detection on the padded, Gaussian-blurred and thresholded version of it. This square was, thus, the extracted DVM from the calibrated OCTA image. The generation process of a DVM is illustrated in FIG. 4.

The process further applied the early treatment of diabetic retinopathy (ETDRS) grid to the DVM generated as described above (FIG. 2, block [5] ) . To drop most of the vessel signals, the map was converted to Lab color space (L, a, b) and all the pixels with

were replaced with white ones to keep only the colored components. The resulting image was next converted to grayscale and skeletonized, and the reference circles in the ETDRS grid were identified through Hough circle transform where the inner reference circle C _inner included a center (x _inner, y _inner) and a radius of r _inner, whereas the outer reference circle C _outer included a center (x _outer, y _outer) = (x _inner, y _inner) and a radius of r _outer. The area within C _inner roughly corresponded to the fovea region, whereas the parafoveal area referred to the ring region between C _inner and C _outer. This process of applying the ETDRS grid to a DVM is illustrated in FIG 5A.

The parafovea was next divided into superior, inferior, nasal and temporal sections by calculating the section boundaries according to whether the studied eye was right (OD) or left (OS) using the recorded parameters of the reference circles in the ETDRS grid. For instance, noting that the coordinates of the top left corner of a DVM are (0, 0) , the end points of the line segment dividing the superior and nasal sections of the right eye are:

and

FIG. 5B demonstrates the sections defined by the ETDRS grid in the left and right eyes.

The process shown in FIG. 2 determined at block [6] the fractal dimension index (FD) based on the DVM from block [4] to quantify the degree of complexity of the retinal microvascular structure in both the superficial capillary plexus (SCP) and the deep capillary plexus (DCP) . The generated DVM is first binarized through adaptive thresholding, and the FD is calculated by applying the box-counting algorithm to the resulting map.

The process automatically delineated retinal vessels in the superficial capillary plexus (FIG. 2, block [7] ) . The process of retinal vessel delineation is illustrated in image A of Figure 6 to image G of Figure 6. The original vasculature map (image A of Figure. 6) was first binarized through adaptive global thresholding, resulting in image B of Figure 6B. Only the fovea and parafovea regions were retained according to the ETDRS grid applied (see FIG. 2, block [5] ) . Next, Sato’s filter was utilized to detect continuous ridges, or the vessels, in the resulting vasculature map (see image C of FIG. 6) . The ridge-filtered image (see image D of FIG. 6) was again binarized through adaptive thresholding based on the median of parafoveal pixel intensity, and small objects, which may be insignificant isolated vessel branches or noise, were removed from the binarized image (see image E of FIG. 6) for better robustness. Finally, all the contours in the resulting map see (image F of FIG. 6) were identified. Each contour corresponded to one delineated retinal vessel as shown in image G of Figure 6.

EXAMPLE 3-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF CPILLARY PERFUSION DENSITY

The process shown in FIG. 2 determined at block [8] the capillary perfusion density index (PDC) based on the generated DVM and the ETDRS grid accompanying it. The map is first converted to grayscale with pixel intensity values ranging from 0 (black) to 255 (white) . Next, PDC is calculated as the average intensity of all the non-delineated-vessel pixels in the considered section of the map. Since brighter pixels (with larger intensity values) often correspond to vascular structures in OCTA images, when a map has a greater calculated PDC, it has denser capillaries.

EXAMPLE 4-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF LARGE VESSELS PERFUSION DENSITY

The process shown in FIG. 2 determined at block [9] the large vessels perfusion density index (PDL) based on the generated DVM in the SCP and its ETDRS grid. For a considered section of the map, PDL is calculated as the proportion of pixels that belong to delineated vessels, or large vessels, in it. When a map has a greater calculated PDL, it has denser large vessels.

EXAMPLE 5-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF VESSEL DISPERSION

The process described in Example 2 was also used to determine the vessel dispersion index (VDisp) (FIG. 2, block [10] and image G of FIG. 6) . In a healthy eye, the parafoveal vessels are more centripetal towards the center of the fovea. Vessel dispersion refers to the degree of centripetalism of the parafoveal vessels. The larger the vessel dispersion is, the less centripetal the vessels are on average. To determine VDisp, the image as shown in image F of FIG. 6 was first padded and cropped so that the parafoveal region was centered. Next, for each delineated retinal vessel V _i (i=1, …, N, where N is the total number of delineated vessels) the ellipse E _i with the same second moments as the region V _i covers was determined. If V _i was perfectly centripetal it was identically orientated as the line segment connecting the centroid of E _i and the center of the fovea. VDisp was thus defined to be the mean of all the angles between the major axis of E _i and the line segment connecting the centroid of E _i and the center of C _inner, which corresponded to the fovea, as identified when applying the ETDRS grid to the DVM as described in FIG. 2, block [5] . This process of VDisp calculation is illustrated in FIG. 7. One example of the determined VDisp’s of individual vessels is shown in Figure 9A. The overall VDisp of an example DVM was the average of all the individual VDisp’s calculated and displayed in the labels as shown, e.g., in FIG. 9A.

The process also separated major and minor retinal vessel segments (FIG. 2, block [11] ) . The process of separating major and minor vessel segments is illustrated in image F of FIG. 6 to image I of FIG. 6. Firstly, the vasculature map with delineated retinal vessels (image F of FIG. 6 and FIG. 2 block [7] ) was skeletonized. The endpoints and branch points in the skeletonized image (see image H of Figure 6) were detected through the algorithm shown in FIG. 8. The identified branch points were then masked from the image and all the contours in the masked image were found. Each identified contour corresponded to one vessel segment. The total number of branch points was denoted as N _bp. The total number of vessel segments N _seg was then equal to 1 + 2N _bp. Further, the geodesic length of a segment was denoted as L _seg and the geodesic length of the major vessel it belonged to was denoted as L _ves. If

the segment was marked as a minor vessel segment. Otherwise, the segment was marked as a major vessel segment.

EXAMPLE 6-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF VESSEL DIAMETER INDEX

The process described was also used to determine the vessel diameter index (VDiam) (FIG. 2, block [12] , image G of FIG. 6, and image H of FIG. 6) . VDiam was defined to be the mean of the ratios between the pixel area and the geodesic length of all the delineated retinal vessels. Denoting S as the side length of the digital vascular map (DVM) , N as the number of delineated vessels and V _i as the i-th delineated vessel in the image, the formula for VDiam was:

An example of the determined VDiam’s of individual vessels is shown in FIG. 9B. The overall VDiam of the DVM was the average of all the labelled individual VDiam’s in the image divided by S.

EXAMPLE 7-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF VESSEL TORTUOSITY

The process described above was further used to determine the vessel tortuosity index (VT) (FIG. 2, block [13] and image I of FIG. 6) . Vessel tortuosity (VT) was quantified as the mean of the ratios between the geodesic length and Euclidean length of the separated vessel segments. Denoting M ₁ as the number of major vessel segments, M ₂ as the number of minor vessel segments, S _i ^major as the i-th major vessel segment and S _i ^minor as the i-th minor vessel segment, the formula for VT was:

Where α=0.8 is a weighting factor. Because minor vessel segments often correspond to secondary vessels less emphasis was laid on them when calculating VT. If vessels in a map were more curved and twisty on average the map had a larger VT. One example of the calculated VT’s of individual vessel segments is shown in FIG. 9C. The overall VT of an example DVM was a weighted average of all the labelled individual VT’s in the map.

EXAMPLE 8-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF THE FOVEAL AVASCULAR ZONE

The process described above was further used to demarcate the foveal avascular zone (FAZ) in a DVM (FIG. 2, block [14] ) . The FAZ was defined to be the capillary-free area within the innermost ring of the parafoveal capillary plexus. The reference circles and lines representing the ETDRS grid in the DVMweare masked with a median filtered version of their corresponding areas in the map. Median filtering was then applied to suppress noise. The masked and denoised image (see, e.g., image B of FIG. 10) was adaptively thresholded and closed (dilation followed by erosion) , followed by small objects removal. Subsequently, bridging (setting 0-valued pixels to 1 if they had two nonzero neighbors that were not connected) and thinning was applied to the resulting map (see, e.g., image C of FIG. 10) . Once again, small isolated objects were removed from the image to improve the robustness of the FAZ identification method.

The maximum polygon in the foveal avascular region, i.e., the polygon that did not cover any of the positive vessel signals was obtained (see, e.g., image D of FIG. 10) . The position (0, 0) corresponded to the top left corner of a DVM. The center of the fovea region was denoted as (x _inner, y _inner) (identified at FIG. 2, block [5] ) and the side length of the DVM was denoted as s. For Θ = 0, 3, 6, 9, …, 357 (i.e., altogether 120 Θ’s ) , all the pixels within the image on the directed line segment from (x _inner, y _inner) to (x _inner+s·sinθ, y _inner+s·cosθ) onsidered. The first pixel with a positive signal, i.e., that is part of a vessel was recorded as one of the endpoints of the FAZ. Half of these endpoints are shown as red dots in image E of FIG. 10. If the values of all the pixels within the image on the directed line segment were zero, the particular Θwas skipped. If less than three searches returned a positive signal, an error was reported. The polygon defined by connecting all the n=120 (or less, if some of the searches did not find any positive pixels) endpoints in the order they were found was the approximated FAZ (see, e.g., image F of Figure 10) . The whole process of demarcating the FAZ is illustrated in FIG. 10.

EXAMPLE 9-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF THE FOVEAL AVASCULAR ZONE AREA INDEX

The process described above was used to determine the FAZ area index (FAI) (FIG. 2, block [15] ) . In clockwise order, the vertices of the FAZ polygon were denoted and demarcated as shown in FIG. 10, as (a ₁, b ₁) , (a ₂, b ₂) , …, (a _n, b _n) . The area of the polygon was calculated using the shoelace formula:

The FAI was next defined to be the FAZ polygon area divided by the pixel area of the DVM. Advantageously, the FAI was comparable across images of different sizes if the DVM was always focused on a 3×3 region.

EXAMPLE 10-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF THE FOVEAL AVASCULAR ZONE PERIMETER INDEX

The process described above was used to determine the FAZ perimeter index (FPI) (FIG. 2, block [16] ) . Following the notations according to FIG. 2, block [15] , the perimeter of the FAZ polygon was calculated using the below formula:

The FPI was then defined to be the FAZ polygon perimeter divided by the side length (in pixels) of the DVM. Advantageously, the FPI was comparable across images of different sizes if the DVM was always focused on a 3×3 region.

EXAMPLE 11-METHODS FOR COMPUTERIZED AUTOMATED OCTA IMAGE ANALYSIS OF THE FOVEAL AVASCULAR ZONE ACIRCULARITY INDEX

The process described was used to determine the FAZ acircularity index (FACI) (FIG. 2, block [17] ) . The FACI quantified the irregularity of the shape of the demarcated FAZ and was calculated using the following formula:

When FACI was equal to one, the FAZ was a perfect circle. Greater FACI values indicated that the FAZ had more irregular shapes.

EXAMPLE 12-MEASUREMENTS OF COMPUTATIONAL RETINAL MICROVASCULAR BIOMARKERS

With the techniques discussed above, different Computational Retinal Microvascular Biomarkers (CRMBs) , including FD over the superficial capillary plexus (SCP) and the deep capillary plexus (DCP) , PDC over the SCP and the DCP, PDL, VDisp, VDiam and VT of the SCP and FAI, FPI and FACI over both SCP and DCP were automatically determined based on OCTA images. To illustrate the usefulness of the proposed computerized automated analytical system, CRMB analysis was carried out for 43 OCTA images from RVO patients and 30 OCTA images of healthy subjects. All OCTA images used in the study were 3 mm × 3 mm scans, were acquired from both SCP and DCP and were analyzed using the automated system of the instant invention. Different CRMBs were determined, representing the degree of retinal abnormalities from both SCP and DCP layers in the macular area. Summary statistics were generated for different CRMBs, and student’s t-tests were performed to compare the magnitude of CRMBs across normal and RVO groups.

fractal dimension index (FDs) of both SCP and DCP layers in eyes with RVO were significantly lower than those of the control eyes (Table 1 and FIG. 11) . Additionally, FAI, FPI and FACI of both layers in eyes with RVO were significantly larger than those of the control eyes.

Vascular Dispersion in eyes with RVO was significantly larger than that of the control eyes. These results showed that the proposed analytical system captured the characteristics clinically important for RVO well. CRMBs were also utilized to study other types of eye diseases to gain insight into the underlying clinical interpretation of the disorders.

TABLE 1. Comparison of CRMBs between normal and RVO groups

EXAMPLE 13-COMPUTATIONAL RETINAL MICROVASCULAR BIOMARKERS

To assess the validity of defined CRMBs, Pearson correlation tests were performed using the systems of the instant invention between ten parafoveal FD phenotypes (SCP and DCP × whole, temporal, superior, nasal and inferior) and the FAZ area provided by the machine-built-in OCTA software. As shown in FIG. 11, all the correlations were strongly positive (ranging from 0.70 to 0.88) and significantly different from zero at a 0.01 adjusted p-value cut-off. This proved the reliability of the computerized automated analytical system of the invention and its calculated CRMBs.

It should be understood that the examples and embodiments described herein are for illustrative purposes only and that various modifications or changes in light thereof will be suggested to persons skilled in the art and are to be included within the spirit and purview of this application. In addition, any elements or limitations of any invention or embodiment thereof disclosed herein can be combined with any and/or all other elements or limitations (individually or in any combination) or any other invention or embodiment thereof disclosed herein, and all such combinations are contemplated with the scope of the invention without limitation thereto.

REFERENCES

Adhi, M., Filho, M.A., Louzada, R.N., Kuehlewein, L., de Carlo, T.E., Baumal, C.R., ... Waheed, N.K. (2016) . Retinal Capillary Network and Foveal Avascular Zone in Eyes with Vein Occlusion and Fellow Eyes Analyzed With Optical Coherence Tomography Angiography. Invest Ophthalmol Vis Sci, 57 (9) , Oct486-494. doi: 10.1167/iovs. 15-18907

Aghamohamadian-Sharbaf, M., Pourreza, H.R., &Banaee, T. (2016) . A Novel Curvature-Based Algorithm for Automatic Grading of Retinal Blood Vessel Tortuosity. IEEE J Biomed Health Inform, 20 (2) , 586-595. doi: 10.1109/jbhi. 2015.2396198

Arrigo, A., Aragona, E., Capone, L., Pierro, L., Romano, F., Bandello, F., &Parodi, M.B. (2018) . Advanced Optical Coherence Tomography Angiography Analysis of Age-related Macular Degeneration Complicated by Onset of Unilateral Choroidal Neovascularization. Am J Ophthalmol, 195, 233-242. doi: 10.1016/j. ajo. 2018.08.001

Arrigo, A., Romano, F., Albertini, G., Aragona, E., Bandello, F., &Battaglia Parodi, M. (2019) . Vascular Patterns in Retinitis Pigmentosa on Swept-Source Optical Coherence Tomography Angiography. J Clin Med, 8 (9) . doi: 10.3390/jcm8091425

Brar, M., Sharma, M., Grewal, S.P.S., &Grewal, D. S. (2019) . Quantification of retinal microvasculature and neurodegeneration changes in branch retinal vein occlusion after resolution of cystoid macular edema on optical coherence tomography angiography. Indian J Ophthalmol, 67 (11) , 1864-1869. doi: 10.4103/ijo. IJO_1554_18

Conrath, J., Valat, O., Giorgi, R., Adel, M., Raccah, D., Meyer, F., &Ridings, B. (2006) . Semi-automated detection of the foveal avascular zone in fluorescein angiograms in diabetes mellitus. Clin Exp Ophthalmol, 34 (2) , 119-123. doi: 10.1111/j. 1442-9071.2006.01138. x

Coscas, F., Glacet-Bernard, A., Miere, A., Caillaux, V., Uzzan, J., Lupidi, M., ... Souied, E. H. (2016) . Optical Coherence Tomography Angiography in Retinal Vein Occlusion: Evaluation of Superficial and Deep Capillary Plexa. Am J Ophthalmol, 161, 160-171. e161-162. doi: 10.1016/j. ajo. 2015.10.008

Costanzo, E., Parravano, M., Gilardi, M., Cavalleri, M., Sacconi, R., Aragona, E., ... Querques, G. (2019) . Microvascular Retinal and Choroidal Changes in Retinal Vein Occlusion Analyzed by Two Different Optical Coherence Tomography Angiography Devices. Ophthalmologica, 242 (1) , 8-15. doi: 10.1159/000496195

Deng, Y., Cai, X., Zhang, S., Su, L., Chen, H., Lin, Y., ... Chi, W. (2019) . Quantitative Analysis of Retinal Microvascular Changes after Conbercept Therapy in Branch Retinal Vein Occlusion Using Optical Coherence Tomography Angiography. Ophthalmologica, 242 (2) , 69-80. doi: 10.1159/000499608

Diaz, M., Novo, J., Cutrin, P., Gomez-Ulla, F., Penedo, M. G., &Ortega, M. (2019) . Automatic segmentation of the foveal avascular zone in ophthalmological OCT-Aimages. PLoS One, 14 (2) , e0212364. doi: 10.1371/journal. pone. 0212364

Dumitrascu, O.M., Demaerschalk, B.M., Valencia Sanchez, C., Almader-Douglas, D., O'Carroll, C.B., Aguilar, M.I., ... Kumar, G. (2018) . Retinal Microvascular Abnormalities as Surrogate Markers of Cerebrovascular Ischemic Disease: A Meta-Analysis. J Stroke Cerebrovasc Dis, 27 (7) , 1960-1968. doi: 10.1016/j. jstrokecerebrovasdis. 2018.02.041

Grisan, E., Foracchia, M., &Ruggeri, A. (2008) . A novel method for the automatic grading of retinal vessel tortuosity. IEEE Trans Med Imaging, 27 (3) , 310-319. doi: 10.1109/tmi. 2007.904657

Hanada, N., Iijima, H., Sakurada, Y., &Imasawa, M. (2012) . Recurrence of macular edema associated with branch retinal vein occlusion after intravitreal bevacizumab. Jpn J Ophthalmol, 56 (2) , 165-174. doi: 10.1007/s10384-011-0113-4

Hao, H., Sasongko, M.B., Wong, T.Y., Che Azemin, M.Z., Aliahmad, B., Hodgson, L., ... Kumar, D.K. (2012) . Does retinal vascular geometry vary with cardiac cycle? Invest Ophthalmol Vis Sci, 53 (9) , 5799-5805. doi: 10.1167/iovs. 11-9326

Hartnett, M. E., Martiniuk, D., Byfield, G., Geisen, P., Zeng, G., &Bautch, V.L. (2008) . Neutralizing VEGF decreases tortuosity and alters endothelial cell division orientation in arterioles and veins in a rat model of ROP: relevance to plus disease. Invest Ophthalmol Vis Sci, 49 (7) , 3107-3114. doi: 10.1167/iovs. 08-1780

Huang, D., Jia, Y., Gao, S.S., Lumbroso, B., &Rispoli, M. (2016) . Optical Coherence Tomography Angiography Using the Optovue Device. Dev Ophthalmol, 56, 6-12. doi: 10.1159/000442770

Kadomoto, S., Muraoka, Y., Ooto, S., Miwa, Y., Iida, Y., Suzuma, K., ... Yoshimura, N. (2018) . EVALUATION OF MACULAR ISCHEMIA IN EYES WITH BRANCH RETINAL VEIN OCCLUSION: An Optical Coherence Tomography Angiography Study. Retina, 38 (2) , 272-282. doi: 10.1097/iae. 0000000000001541

Khodabandeh, A., Shahraki, K., Roohipoor, R., Riazi-Esfahani, H., Yaseri, M., Faghihi, H., &Bazvand, F. (2018) . Quantitative measurement of vascular density and flow using optical coherence tomography angiography (OCTA) in patients with central retinal vein occlusion: Can OCTA help in distinguishing ischemic from non-ischemic type? Int J Retina Vitreous, 4, 47. doi: 10.1186/s40942-018-0152-9

Kulikov, A.N., Maltsev, D.S., &Burnasheva, M.A. (2018) . Improved analysis of foveal avascular zone area with optical coherence tomography angiography. Graefes Arch Clin Exp Ophthalmol, 256 (12) , 2293-2299. doi: 10.1007/s00417-018-4139-x

Lee, H., Lee, M., Chung, H., &Kim, H.C. (2018) . QUANTIFICATION OF RETINAL VESSEL TORTUOSITY IN DIABETIC RETINOPATHY USING OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY. Retina, 38 (5) , 976-985. doi: 10.1097/iae. 0000000000001618

Lee, H.E., Wang, Y., Fayed, A.E., &Fawzi, A.A. (2019) . Exploring the relationship between collaterals and vessel density in retinal vein occlusions using optical coherence tomography angiography. PLoS One, 14 (7) , e0215790. doi: 10.1371/journal. pone. 0215790

McIntosh, R.L., Rogers, S.L., Lim, L., Cheung, N., Wang, J.J., Mitchell, P., ... Wong, T.Y. (2010) . Natural history of central retinal vein occlusion: an evidence-based systematic review. Ophthalmology, 117 (6) , 1113-1123. e1115. doi: 10.1016/j. ophtha. 2010.01.060

Murray, C. D. (1926a) . The Physiological Principle of Minimum Work: I. The Vascular System and the Cost of Blood Volume. Proc Natl Acad Sci U S A, 12 (3) , 207-214. doi: 10.1073/pnas. 12.3.207

Murray, C.D. (1926b) . The Physiological Principle of Minimum Work: II. Oxygen Exchange in Capillaries. Proc Natl Acad Sci U S A, 12 (5) , 299-304. doi: 10.1073/pnas. 12.5.299

Newman, A.R., Andrew, N.H., &Casson, R.J. (2017) . Review of paediatric retinal microvascular changes as a predictor of cardiovascular disease. Clin Exp Ophthalmol, 45 (1) , 33-44. doi: 10.1111/ceo. 12773

Papadia, M., Misteli, M., Jeannin, B., &Herbort, C.P. (2014) . The influence of anti-VEGF therapy on present day management of macular edema due to BRVO and CRVO: a longitudinal analysis on visual function, injection time interval and complications. Int Ophthalmol, 34 (6) , 1193-1201. doi: 10.1007/s10792-014-0002-1

Qian, T., Zhao, M., &Xu, X. (2017) . Comparison between anti-VEGF therapy and corticosteroid or laser therapy for macular oedema secondary to retinal vein occlusion: A meta-analysis. J Clin Pharm Ther, 42 (5) , 519-529. doi: 10.1111/jcpt. 12551

Rispoli, M., Savastano, M.C., &Lumbroso, B. (2015) . CAPILLARY NETWORK ANOMALIES IN BRANCH RETINAL VEIN OCCLUSION ON OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY. Retina, 35 (11) , 2332-2338. doi: 10.1097/iae. 0000000000000845

Rogers, S., McIntosh, R.L., Cheung, N., Lim, L., Wang, J.J., Mitchell, P., ... Wong, T.Y. (2010) . The prevalence of retinal vein occlusion: pooled data from population studies from the United States, Europe, Asia, and Australia. Ophthalmology, 117 (2) , 313-319. e311. doi: 10.1016/j. ophtha. 2009.07.017

Rogers, S.L., McIntosh, R.L., Lim, L., Mitchell, P., Cheung, N., Kowalski, J. W., ... Wong, T. Y. (2010) . Natural history of branch retinal vein occlusion: an evidence-based systematic review. Ophthalmology, 117 (6) , 1094-1101. e1095. doi: 10.1016/j. ophtha. 2010.01.058

Sasongko, M.B., Hodgson, L.A., Wong, T.Y., Kawasaki, R., Cheung, C.Y., Hsu, W., ... Wang, J.J. (2012) . Correlation and reproducibility of retinal vascular geometric measurements for stereoscopic retinal images of the same eyes. Ophthalmic Epidemiol, 19 (5) , 322-327. doi: 10.3109/09286586.2012.702258

Sasongko, M.B., Wang, J.J., Donaghue, K.C., Cheung, N., Benitez-Aguirre, P., Jenkins, A., ... Wong, T.Y. (2010) . Alterations in retinal microvascular geometry in young type 1 diabetes. Diabetes Care, 33 (6) , 1331-1336. doi: 10.2337/dc10-0055

Seknazi, D., Coscas, F., Sellam, A., Rouimi, F., Coscas, G., Souied, E.H., &Glacet-Bernard, A. (2018) . OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY IN RETINAL VEIN OCCLUSION: Correlations Between Macular Vascular Density, Visual Acuity, and Peripheral Nonperfusion Area on Fluorescein Angiography. Retina, 38 (8) , 1562-1570. doi: 10.1097/iae. 0000000000001737

Spaide, R. F. (2017) . MICROVASCULAR FLOW ABNORMALITIES ASSOCIATED WITH RETINAL VASCULITIS: A Potential of Mechanism of Retinal Injury. Retina, 37 (6) , 1034-1042. doi: 10.1097/iae. 0000000000001359

Spaide, R.F., Klancnik, J.M., Jr., &Cooney, M.J. (2015) . Retinal vascular layers imaged by fluorescein angiography and optical coherence tomography angiography. JAMA Ophthalmol, 133 (1) , 45-50. doi: 10.1001/jamaophthalmol. 2014.3616

Suzuki, N., Hirano, Y., Yoshida, M., Tomiyasu, T., Uemura, A., Yasukawa, T., &Ogura, Y. (2016) . Microvascular Abnormalities on Optical Coherence Tomography Angiography in Macular Edema Associated With Branch Retinal Vein Occlusion. Am J Ophthalmol, 161, 126-132. e121. doi: 10.1016/j. ajo. 2015.09.038

Winegarner, A., Wakabayashi, T., Fukushima, Y., Sato, T., Hara-Ueno, C., Busch, C., ... Nishida, K. (2018) . Changes in Retinal Microvasculature and Visual Acuity After Antivascular Endothelial Growth Factor Therapy in Retinal Vein Occlusion. Invest Ophthalmol Vis Sci, 59 (7) , 2708-2716. doi: 10.1167/iovs. 17-23437

Witt, N., Wong, T.Y., Hughes, A.D., Chaturvedi, N., Klein, B.E., Evans, R., ... Klein, R. (2006) . Abnormalities of retinal microvascular structure and risk of mortality from ischemic heart disease and stroke. Hypertension, 47 (5) , 975-981. doi: 10.1161/01. HYP. 0000216717.72048.6c

Yamada, R., Nishida, A., Shimozono, M., Kameda, T., Miyamoto, N., Mandai, M., &Kurimoto, Y. (2015) . Predictive factors for recurrence of macular edema after successful intravitreal bevacizumab therapy in branch retinal vein occlusion. Jpn J Ophthalmol, 59 (6) , 389-393. doi: 10.1007/s10384-015-0412-2

Yoo, S.J., Kim, J.H., Lee, T.G., Kim, J.W., Cho, S.W., &Han, J.I. (2017) . Natural Short-term Course of Recurrent Macular Edema Following Intravitreal Bevacizumab Therapy in Branch Retinal Vein Occlusion. Korean J Ophthalmol, 31 (2) , 95-101. doi: 10.3341/kjo. 2017.31.2.95

Youm, D.J., Ha, M.M., Chang, Y., &Song, S.J. (2012) . Retinal vessel caliber and risk factors for branch retinal vein occlusion. Curr Eye Res, 37 (4) , 334-338. doi: 10.3109/02713683.2011.629070

Lemmens, Sophie, et al. "Systematic review on fractal dimension of the retinal vasculature in neurodegeneration and stroke: assessment of a potential biomarker. " Frontiers in neuroscience 14 (2020) : 16

Claims

A method for automated quantitative assessment of retinal microvasculature and generation of a retinal digital vascular map comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data; and

generating a retinal digital vascular map (DVM) .
The method according to claim 1, wherein the pre-processing comprises:

applying an Early Treatment Diabetic Retinopathy Study grid to the DVM;

calculating a fractal dimension index; and

demarcating retinal blood vessels.
The method according to claim 2 wherein the pre-processing further comprises:

calculating a blood vessel dispersion index;

separating major and minor blood vessel segments;

calculating a blood vessel diameter index;

calculating a blood vessel tortuosity index;

demarcating a foveal avascular zone;

calculating a foveal avascular zone area index; and

calculating a foveal avascular zone acircularity index.
A method of extracting computational retinal microvascular biomarkers (CRMBs) from a DVM generated using the method of claim 1.
The method according to claim 4, wherein the CRMB is selected from fractal dimension index, capillary perfusion density index, large vessels perfusion density index, blood vessel dispersion index, blood vessel diameter index, blood vessel tortuosity index, foveal avascular zone area index, foveal avascular zone perimeter index and foveal avascular zone acircularity index.
A method for automated quantification of computational retinal microvascular biomarkers from a retinal superficial capillary plexus of a subject suspected to suffer from a retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from a retinal disease, the method comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

generating a retinal digital vascular map (DVM) of the superficial capillary plexus;

providing a retinal DVM of a normal subject;

quantifying at least one CRMB in the DVM of the subject and the DVM of the normal subject; and

treating the subject for a retinal disease if the quantity of the at least one CRMB in the subject exceeds the quantity of said CRMB in the normal subject.
The method according to claim 6, wherein the pre-processing comprises: applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and wherein the quantifying at least one CRMB comprises:

quantifying a fractal dimension index;

quantifying a capillary perfusion density index;

quantifying a large vessels perfusion density index;

quantifying a blood vessel dispersion index;

quantifying a blood vessel diameter index;

quantifying a blood vessel tortuosity index;

quantifying a foveal avascular zone area index;

quantifying a foveal avascular zone perimeter index and/or

quantifying a foveal avascular zone acircularity index.
The method according to claim 7, wherein the retinal disease is selected from diabetic retinopathy, retinal vein occlusion, and age-related macular degeneration and the treatment comprises administering an antibody to a vascular growth factor, an aptamer to a vascular growth factor, an anti-angiogenic steroid, an anti-oxidant supplementation, a laser photocoagulation therapy, a transpupillary thermotherapy, and/or performing an eye surgery.
A method for automated quantification of computational retinal microvascular biomarkers from a retinal deep capillary plexus of a subject suspected to suffer from a retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from a retinal disease, the method comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

generating a retinal digital vascular map (DVM) of the deep capillary plexus;

providing a retinal DVM of a deep capillary plexus of a normal subject;

quantifying at least one CRMB in the DVM of the subject and the DVM of the normal subject; and

treating the subject for a retinal disease if the quantity of the at least one CRMB in the subject exceeds the quantity of said CRMB in the normal subject.
The method according to claim 9, wherein the pre-processing comprises applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and the quantifying at least one CRMB comprises:

quantifying a fractal dimension index;

quantifying a capillary perfusion density index;

quantifying a large vessels perfusion density index;

quantifying a blood vessel dispersion index;

quantifying a blood vessel diameter index;

quantifying a blood vessel tortuosity index;

quantifying a foveal avascular zone area index;

quantifying a foveal avascular zone perimeter index and/or

quantifying a foveal avascular zone acircularity index.
The method according to claim 9, wherein the retinal disease is selected from diabetic retinopathy, retinal vein occlusion, and age-related macular degeneration and the treatment comprises administering an antibody to a vascular growth factor, an aptamer to a vascular growth factor, an anti-angiogenic steroid, an anti-oxidant supplementation, a laser photocoagulation therapy, a transpupillary thermotherapy, and/or performing an eye surgery.
A method for automated quantification of computational retinal microvascular biomarkers from a retinal superficial capillary plexus of a subject suspected to suffer from an extra-retinal disease and treatment of the subject if the CRMB quantification indicates that the subject suffers from an extra-retinal disease, the method comprising:

recording optical coherence tomography angiography (OCTA) image data;

pre-processing OCTA image data;

generating a retinal digital vascular map (DVM) of the superficial capillary plexus;

providing a retinal DVM of a normal subject;

quantifying at least one CRMB in the DVM of the subject and the DVM of the normal subject; and

treating the subject for an extra-retinal disease if the quantity of the at least one CRMB in the subject exceeds the quantity of said CRMB in the normal subject.
The method according to claim 12, wherein the pre-processing comprises applying an Early Treatment Diabetic Retinopathy Study grid to the DVM and the quantifying at least one CRMB comprises:

quantifying a fractal dimension index;

quantifying a capillary perfusion density index;

quantifying a large vessels perfusion density index;

quantifying a blood vessel dispersion index;

quantifying a blood vessel diameter index;

quantifying a blood vessel tortuosity index;

quantifying a foveal avascular zone area index;

quantifying a foveal avascular zone perimeter index and/or

quantifying a foveal avascular zone acircularity index.
The method according to claim 13, wherein the extra-retinal disease is selected from cardiovascular disease, systemic arterial hypertension, sickle cell anemia, anemia, pre- eclampsia, arteritis, arterial macroaneurysm, diabetes mellitus, and/or a hypercoagulation disorder and the treatment comprises administering an anti-hypertensive drug, a cholesterol-modifying drug, aspirin, a beta receptor blocker, a calcium channel blocker, nitroglycerin, an angiotensin-converting enzyme inhibitor, an angiotensin II receptor blocker, a blood thinning drug; an iron supplementation, a steroid, an antidiabetic drug, an endovascular stent surgery and/or a vascular bypass surgery.
An automated analysis system for optical coherence tomography angiography (OCTA) image analysis, comprising: a storage, configured to store one or more executable instructions; a processor, configured to implement the steps of the method of any one of claims 1-14 by executing the one or more executable instructions.
A computer-readable storage medium, which contains a computer program, and the computer program can be executed by a processor to implement the steps of the method of any one of claims 1-14.