Artificial intelligence algorithm to assist in the diagnosis of ocular diseases using optical images

Abstract

Systems and methods are provided for identifying an ocular affliction (e.g., ocular surface squamous neoplasia (OSSN)) via ocular images (e.g., high resolution optical coherence tomography (OCT) (HR-OCT), slit lamp photographs, or other photographs) using deep learning (DL). A trained DL model can be used to determine from an ocular image whether a patient has an ocular affliction or not. The DL model can be trained using ocular image data with assigned annotations by experts indicating normal tissues and those with an ocular affliction (e.g., OSSN and other conditions).

Description

[0001] DESCRIPTION

[0002] ARTIFICIAL INTELLIGENCE ALGORITHM TO ASSIST IN THE DIAGNOSIS OF OCULAR DISEASES USING OPTICAL IMAGES

[0003] CROSS-REFERENCE TO RELATED APPLICATION

[0004] This application claims the benefit of U.S. Provisional Application Serial No. 63 / 627,578, filed January 31 , 2024, the disclosure of which is hereby incorporated by reference in its entirety, including all figures, tables, and drawings.

[0005] BACKGROUND

[0006] Ocular surface squamous neoplasia (OSSN) is the most common nonpigmented ocular surface malignancy. It represents an abnormal growth of squamous epithelial tissue on the cornea or conjunctiva, comprising a histological range from mild epithelial dysplasia to invasive squamous cell carcinoma. The incidence of OSSN ranges from 0.03-1.9 per 100,000 persons / year in the United States and Australia to 1.6-3.4 per 100,000 persons / year in sub- Saharan Africa. Ocular surface cancer has significant morbidity, leading to visual compromise or limbal stem cell deficiency and can cause mortality if there is tumor invasion into the sclera or orbit and subsequent metastasis. OSSN is typically diagnosed through excisional biopsy (and histopathological analysis of the lesion), which involves complications such as scarring, limbal stem cell deficiency, and seeding of tumor. Also, OSSN and many other ocular afflictions can be challenging to diagnose, especially by non-experts.

[0007] BRIEF SUMMARY

[0008] Embodiments of the subject invention provide novel and advantageous systems and methods for identifying diseases and / or tumors affecting the eye (and / or in the eye), and / or differentiating ocular surface squamous neoplasia (OSSN) from other pathologies, via ocular images (e.g., high resolution optical coherence tomography (OCT) (HR-OCT) or slit lamp images) using deep learning (DL) and / or machine learning (ML). A trained DL model can be used to determine from an ocular image (e.g., HR-OCT or slit lamp image) whether a patient has a particular disease or tumor or not. The DL model can be trained using ocular image (e.g., HR- OCT, slit lamp image, or other photographic image) data with assigned annotations by experts indicating normal tissues and those with a particular disease or tumor (e.g., OSSN, melanoma, lymphoma, and / or other ocular conditions (e.g., cataracts, ulcers, dry eye, pterygia, corneal or other ocular afflictions)).

[0009] In an embodiment, a system for DL-based diagnosis of an ocular affliction (e.g., an ocular disease or tumor, such as OSSN, melanoma, lymphoma, and / or other ocular conditions (e.g., cataracts, ulcers, dry eye, pterygia, corneal or other ocular afflictions)) can comprise: a processor; and a machine-readable medium in operable communication with the processor and having instructions stored thereon that, when executed by the processor, perform the following steps: a) receiving first data of an image (e.g., an HR-OCT image or an ocular image (including a photograph taken from a slit lamp, cell phone, or any photographic device)) of an eye of a patient; b) using a convolutional neural network (CNN) to perform segmentation (e.g., epithelial (segmentation) or using any other key feature(s)) on the first data; c) using the CNN to determine second data of a first characteristic / property (e.g., the epithelial thickness or utilizing any other key feature(s)) of the eye, the second data comprising at least one of a peak of the first characteristic of the eye and an average of the first characteristic of the eye; and d) comparing the second data to a trained DL model to make a determination of whether the patient has the ocular affliction or not. The instructions when executed can further comprise, before step d), training a DL model using a large number (e.g., at least 20,000, at least 50,000, at least 60,000, at least 70,000, at least 80,000) of ocular images (e.g., HR-OCT images, slit lamp images, or any other photographic images) with known diagnostic information to obtain the trained DL model. The known diagnostic information can comprise, for example, annotations from expert graders. The training of the DL model can comprise using an encoder and a custom head classifier. The custom head classifier can comprise a vision transformer supervised model. The second data can comprise both the peak of the first characteristic (e.g., epithelial thickness or any other key feature(s)) of the eye and the average of the first characteristic (e.g., epithelial thickness or any other key feature(s)) of the eye. The determination of whether the patient has the ocular affliction or not can be based on if the second data is within a predetermined threshold from one or more benchmark values from the trained DL model. The determination of whether the patient has the ocular affliction or not can comprise using a vision transformer. The system can further comprise a display in operable communication with the processor and / or the machine-readable medium. The instructions when executed can further comprise: e) displaying, on the display, the determination of whether the patient has the ocular affliction or not. Any or all ocular images used (e.g., for the first data and / or the training data) can be anterior segment ocular image(s).

[0010] In another embodiment, a method for DL-based diagnosis of an ocular affliction can comprise: a) receiving (e.g., by a processor) first data of an image (e.g., an HR-OCT image or an ocular image (including a photograph taken from a slit lamp, cell phone, or any photographic device)) of an eye of a patient; b) using (e.g., by the processor) a CNN to perform segmentation (e.g., epithelial (segmentation) or using any other key feature(s)) on the first data; c) using (e.g., by the processor) the CNN to determine second data of a first characteristic (e.g., the epithelial thickness or utilizing any other key feature(s)) of the eye, the second data comprising at least one of a peak of the first characteristic of the eye and an average of the first characteristic of the eye; and d) comparing (e.g., by the processor) the second data to a trained DL model to make a determination of whether the patient has the ocular affliction or not. The method can further comprise, before step d), training (e.g., by the processor) a DL model using a large number (e.g., at least 20,000, at least 50,000, at least 60,000, at least 70,000, at least 80,000) of ocular images (e.g., HR-OCT images, slit lamp images, or other photographic images) with known diagnostic information to obtain the trained DL model. The known diagnostic information can comprise annotations from expert graders. The training of the DL model can comprise using an encoder and a custom head classifier. The custom head classifier can comprise a vision transformer supervised model. The second data can comprise both the peak of the first characteristic (e.g., epithelial thickness or any other key feature(s)) of the eye and the average of the first characteristic (e.g., epithelial thickness or any other key feature(s)) of the eye. The determination of whether the patient has the ocular affliction or not can be based on if the second data is within a predetermined threshold from a benchmark value from the trained DL model. The determination of whether the patient has the ocular affliction or not can comprise using (e.g., by the processor) a vision transformer. The method can further comprise: e) displaying (e.g., by the processor) the determination of whether the patient has the ocular affliction or not (e.g., on a display, such as a display in operable communication with the processor). Any or all ocular images used (e.g., for the first data and / or the training data) can be anterior segment ocular image(s).

[0011] BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 shows a flow diagram depicting the data collection and data analysis process to establish a diagnosis (e.g., of ocular surface squamous neoplasia (OSSN)) (e.g., using high resolution optical coherence tomography (OCT) (HR-OCT) scans and slit lamp photographs). Though Figure 1 lists “Slit Lamp Photo”, this is for exemplary' purposes only; any type of image can be used as the starting point. Additionally, while Figure 1 shows the Decision to be either “OSSN” or “Non OSSN”, this is for exemplary purposes as well; the model can evaluate for the presence of any ocular affliction.

[0013] Figures 2A-2D show example clinical photos and associated HR-OCT images of OSSN and subclinical OSSN identified on HR-OCT. Figure 2A shows a slit lamp image of obvious OSSN located on the temporal conjunctiva of the left eye with extension on to the cornea at 3-4 o’clock. The eye is injected with multiple associated feeder vessels. Figure 2B shows a corresponding HR-OCT of the obvious OSSN taken at the temporal limbus. There is epithelial thickening, epithelial hyperreflectivity, and an abrupt transition from normal to abnormal epithelium. Figure 2C shows a slit lamp image of subtle, subclinical OSSN located on the nasal conjunctiva of the right eye. The eye is quiet and not obvious cancer noted on this image. Figure 2D shows the corresponding HR-OCT of subclinical OSSN taken at the nasal limbus. The HR- OCT was able to clearly demonstrate the features of OSSN including epithelial thickening, epithelial hyperreflectivity, and an abrupt transition from normal to abnormal epithelium (box).

[0014] Figure 3 shows HR-OCT images depicting three well-described characteristics of OSSN: (1) a thickened epithelium; (2) a hyperreflective epithelium; and (3) an abrupt (vertical) transition from normal to abnormal epithelium (box).

[0015] Figures 4A and 4B show example identification of two cases of cancerous lesions found on HR-OCT that were correctly diagnosed by a deep learning (DL) model of systems and methods of embodiments of the subject invention as OSSN and not by expert graders.

[0016] Figure 5 shows a diagram depicting the overall steps to build the DL methodology in the DL model of systems and method of embodiments of the subject invention. Figure 6 shows a table of demographic characteristics of a patient sample in a test dataset.

[0017] Figure 7 shows a table of diagnostic performance of a DL model, according to an embodiment of the subject invention, compared with expert clinicians to assess the presence of OSSN using HR-OCT scans.

[0018] Figure 8 shows a table of diagnostic performance of a DL model, according to an embodiment of the subject invention, using slit lamp photographs to assess the presence of OSSN.

[0019] DETAILED DESCRIPTION

[0020] Embodiments of the subject invention provide novel and advantageous systems and methods for identifying diseases and / or tumors in the eye, and / or differentiating ocular surface squamous neoplasia (OSSN) from other pathologies, via ocular images (e.g., high resolution optical coherence tomography (OCT) (HR-OCT), slit lamp images, or any other photograph of the eye) using deep learning (DL) and / or machine learning (ML). A trained DL model can be used to determine from an ocular image (e.g., HR-OCT, slit lamp image, or any other photograph) whether a patient has a particular disease or tumor (e.g., OSSN or other ocular affliction) or not. The DL model can be trained using ocular image (e.g., HR-OCT, slit lamp image, or other photographic image) data with assigned annotations by experts indicating normal tissues and those with a particular disease or tumor (e.g., OSSN, melanoma, lymphoma, and / or other conditions (e.g., cataracts, ulcers, dry eye, pterygia, corneal or other ocular afflictions)).

[0021] HR-OCT has a sensitivity and specificity of 100% in differentiating OSSN from pterygium (cutoff epithelial thickness of 120 micrometers (pm)). However, most clinicians are uncomfortable and / or inexperienced in reading HR-OCT and cannot take advantage of this technology. Embodiments of the subject invention advantageously utilize DL to provide diagnostic interpretation of HR-OCT to help enhance patient care.

[0022] OCT is a non-invasive, in-office imaging technique to examine the anterior and posterior segment of the eye. Spectral domain OCT provides a cross-sectional view of the retina and its images can be used to diagnose and monitor conditions such as glaucoma, macular degeneration, and macular edema. OCT can be applied to the anterior segment of the eye, and HR-OCT can use a longer wavelength of light to enhance tissue penetration, allowing for better characterization of ocular surface lesions.

[0023] HR-OCT can help to differentiate OSSN from benign lesions with high precision but in the related art requires expert interpretation. OSSN can be identified on HR-OCT (see Figures 2A-2D) using three well-described characteristics: (1) a thickened epithelium; (2) a hyperreflective epithelium; and (3) an abrupt (vertical) transition from normal to abnormal epithelium. A sensitivity and specificity of 100% can be obtained in differentiating a diagnosis of OSSN from pterygium using a cutoff epithelial thickness of 120 pm using the commercially available Optovue HR-OCT device (Optovue, Fremont, CA) (see also Nanji et al., High- Resolution Optical Coherence Tomography as an Adjunctive Tool in the Diagnosis of Corneal and Conjunctival Pathology, Ocul Surf, 2015, 13(3): p. 226-35; which is hereby incorporated by reference herein in its entirety). However, similar characteristics have been noted in various HR- OCT instruments, including Topcon 3D OCT-1 Maestro (3D OCT-1 Maestro, Topcon, Tokyo) and Heidelberg Spectral Domain OCT (Heidelberg Engineering, Heidelberg, Germany). The utility of HR-OCT has been demonstrated by its ability to help identify cases of OSSN that present with subclinical features not easily identifiable on slit lamp exam (see also Figures 2C and 2D) (see also Tran et al., Utility of high-resolution anterior segment optical coherence tomography in the diagnosis and management of sub-clinical ocular surface squamous neoplasia, Eye Vis (Lond), 2019, 6: p. 27; which is hereby incorporated by reference herein in its entirety). In clinical settings, it is routinely used to monitor treatment responses to both medical and surgical modalities, and has been shown to predict the surgical margins of OSSN intraoperatively.

[0024] Several major gaps remain in the diagnostic paradigm of corneal and conjunctival lesions. One is the need for a tissue biopsy to definitively differentiate neoplastic from non-neoplastic lesions consistently. Tissue biopsies are invasive, costly, time-consuming, require a skilled ocular pathologist for interpretation, and may be associated with complications including seeding of tumor, scarring, infection, and dry eye. When lesions are subtle, a biopsy may miss the diagnosis completely with a false negative. There are many benefits to early identification of disease. Early recognition and treatment of OSSN, other tumors, or other eye conditions can improve visual prognosis and reduce the risk of complications. Further, the imaging can evaluate the entire ocular surface, something not possible with biopsy alone. HR-OCT has bridged this gap as an “optical biopsy”, as it has been shown to noninvasively improve precision in diagnosing ocular surface lesions. However, a gap remains in that the clinical utility of HR-OCT still depends on an expert interpreting the images. The automated analysis of HR-OCT images, as provided by embodiments of the subject invention, can allow for widespread adoption of HR-OCT technology and serve a larger patient population.

[0025] Artificial intelligence (Al) techniques, such DL, can be used to assist in the interpretation of imaging to improve the diagnosis of several common ophthalmic conditions such as glaucoma, diabetic retinopathy, corneal diseases, and macular degeneration. Supervised DL models, such as convolutional neural networks (CNN), are validated tools that are useful in medical image understanding such as image classification tasks. CNN-based approaches can be applied to the detection and progression of ocular diseases (e.g., glaucoma) with high sensitivities and specificities. In relation to the ocular surface, CNNs can be trained with slit lamp photographs to differentiate between active comeal ulcers and healed scars as well as between bacterial and fungal keratitis with high accuracy.

[0026] Embodiments of the subject invention provide systems and methods that use a DL algorithm to assist in the diagnosis of OSSN using HR-OCT images. Specifically, the DL model aims to distinguish OSSN from other common clinically confounding lesions such as pterygium, pinguecula, melanoma, nevus, and lymphoma (e.g., using a CNN). The DL model can achieve diagnostic performance comparable to that of expert clinician interpretation of HR-OCT for biopsy -proven lesions, the reference standard.

[0027] Figure 1 shows a flow chart of an algorithm that can be used with systems and methods of embodiments of the subject invention (for any ocular affliction), and Figure 5 shows another view of a DL method that can be used with systems and methods of embodiments of the subject invention. Referring to Figures 1 and 5, HR-OCT scans (or any other photographic image) can be used to train an encoder / vision transformer, followed by a custom head classifier. In particular, the DL model can be constructed in a two-step procedure, where unlabeled data (HR- OCT scans or any photographic image) can be used to train a masked autoencoder to reduce the images into a low-dimensional latent space, followed by randomly selected images for labeling being used as input for fine-tuning of the binary classifier. After the DL model is trained, it can be used for classifying an HR-OCT image (or any anterior segment photograph) as having OSSN or not (“non-OSSN”), or identify any ocular condition. For example, a CNN (with epithelial segmentation) can be used on an HR-OCT image, where peak and average thickness are obtained from the HR-OCT image. This data is then compared to the trained DL model, and a decision is made as to whether the patient from the HR-OCT scan has OSSN or not. The decision can be based on, for example, if the peak and / or average thickness are within a predetermined threshold from benchmark values from the trained DL model.

[0028] In some embodiments, an encoder / decoder can be used in the training, and then an encoder and vision transformer can be used as part of the CNN and / or determination of whether the patient from HR-OCT scan has an ocular affliction (e.g., OSSN or any ocular affliction) or not (see also Figure 5).

[0029] In another embodiment, slit lamp or any photographic image can be used instead of HR- OCT scans to diagnose ocular conditions. The slit lamp photo(s) can be run through a CNN for quality assessment, followed by an encoder / vision transformer and a custom head classifier. The results can be compared with a trained DL model to make a decision as to whether the patient from the slit lamp photo has an ocular affliction (e.g., OSSN or other affliction (e.g., melanoma, lymphoma, cataracts, ulcers, dry eye, pterygia, corneal or any ocular afflictions)) or not (see also the results in Figure 8).

[0030] Embodiments of the subject invention provide diagnostic tools to differentiate OSSN from other clinically confounding lesions, such as pterygium, pinguecula, and other ocular surface malignancies / lesions using HR-OCT imaging and DL methodology. The classification tool uses individual HR-OCT B-scans to assign a diagnosis, and it performs with high accuracy in identifying OSSN (see also Example 1). The algorithm can achieve higher diagnostic sensitivity, specificity, and accuracy for two-class classification of OSSN versus pterygium and pinguecula in comparison to expert graders (who were found to have a poor inter-grader agreement score (kappa = 15%), in the reading of HR-OCT images alone, in Example 1).

[0031] While HR-OCT has proven efficacy as a diagnostic instrument for the clinical assessment of ocular surface lesions, non-objective observer metrics for image assessment can allow for disagreement in diagnostic interpretations. Similar discrepancies have been noted in other fields within ophthalmology. For example, the cup-to-disk ratio, a standard clinical assessment for glaucomatous damage, exhibits poor inter and intra observer agreement between expert clinicians. When HR-OCT images are used to assess for OSSN, characteristics like epithelial thickening, hyper-reflectivity, and an abrupt transition zone are noted. However, these features are not always quantitatively evaluated and standardized in terms of thickness, degree of hyperreflectivity, or precise measurement of the black and white gradient or contrast for the definition of OSSN versus non-OSSN lesions. The assessment of these features largely depends on the clinician’s expertise and subjective impression, likely contributing to the inter-observer variability noted in Example 1, much like what has been observed in optic nerve imaging research. These results highlight the need for automated assessment of HR-OCT images that utilize objective criteria in establishing clinical diagnoses.

[0032] The findings in Example 1 support the premise that artificial intelligence can greatly improve the diagnosis of ocular surface tumors, improve timely detection of lesions and subclinical disease, and improve screening. Ophthalmology is a uniquely visual medical specialty, and hence lends itself to the incorporation of DL tools to help guide diagnosis and treatment decision-making. DL techniques have been helpful with respect to fundus photographs to identify various retinal pathologies including dry age-related macular degeneration, neovascular age related macular degeneration, glaucomatous damage, and diabetic retinopathy, among others. In addition, DL-based screening tools have received Food and Drug Administration (FDA) approval in the early monitoring of diabetic retinopathy and have been proposed in conditions such as glaucoma, with the aim of increasing access to care and reducing healthcare disparities. Similarly, embodiments of the subject invention can improve care by providing comprehensive ophthalmologists the opportunity to conduct screenings or guide preliminary assessments before deciding whether a referral or biopsy is required. This approach can enhance the efficiency of preliminary screenings and improve the quality of referrals to specialized ophthalmologists with respect to ocular surface tumors and other ocular afflictions as well.

[0033] DL classification tools have not previously been applied to OCT or slit lamp photographs for the detection of OSSN, which can be clinically challenging to diagnose and may masquerade as many other inflammatory, neoplastic, and non-neoplastic lesions. In addition, existing DL classification tools in ocular oncology are limited due to reliance on exclusively supervised DL models.

[0034] Within DL, supervised and unsupervised neural network sub classes allow for different approaches in training models for image classification tasks. Embodiments of the subject invention can utilize a novel methodology in the development of the DL model, involving pretraining the supervised algorithm with a large database of HR-OCT scans (e.g., scans collected from 2015-2021) with an unsupervised autoencoder. Autoencoders are state of the art DL technology that have been applied to various classification tasks in medicine, including ophthalmology. They are effective at reducing noise in images and augmenting the potential of DL models to work with smaller, labeled datasets - a problem encountered by many who wish to train supervised DL models. Due to low inter-grader agreement scores between diagnostic labels, a foundational unsupervised framework allowed for the model to learn from a more objective reference standard, as the machine learns how to represent the dataset without external input. To the best of our knowledge, these methods have never been applied to the ocular surface.

[0035] The algorithm can incorporate a retrospective cohort of biopsy -proven lesions that have HR-OCT scans. Such samples may be subject to selection bias, as lesions that are biopsied are typically those that either could not be diagnosed clinically or were large and hence required surgical excision over medical treatment. While each lesion has biopsy confirmation, each OCT B-scan may not necessarily be able to be associated with the site of biopsy in each patient and a diagnosis. While the model was assessed on biopsy proven lesions, each HR-OCT B-scan on the testing dataset required subjective interpretation by a human clinician and cannot be an objective ground truth for assessment. Additionally, all HR-OCT exams were administered at one clinical site and on one device model and may not represent a generalizable sample. Challenges with the application of DL to anterior segment HR-OCT include the variability between B-scans for lesions. B-scans are often taken at various angles relative to the lesion, include different anatomical landmarks - such as the including cornea, limbus, bulbar and palpebral conjunctiva - and are subject to artifacts (i.e., thickness of the lesion causing posterior shadowing and reduced visibility).

[0036] Despite the limitations, the development of the models of embodiments of the subject invention serves as a proof of concept to illustrate the promising applications for DL with ocular imaging (e.g., HR-OCT imaging, slit lamp imaging, or other photographic imaging) for the diagnosis of ocular surface lesions and other ocular afflictions. Embodiments can classify OSSN lesions with HR-OCT scans after developing and applying a DL classification model. After training and testing several models, the algorithm was able to diagnose cancer with a higher sensitivity and specificity than expert graders using HR-OCT images alone (see Example 1).

[0037] Embodiments of the subject invention can also include DL CNNs for eye afflictions that do not use the epithelial thickness at all (e.g., cataracts). That is, the DL CNN can pertain to other anterior segment afflictions that can have a methodology that does not use epithelial segmentation, such as those using OCT or any other ocular photograph.

[0038] Embodiments of the subject invention can use a vision transformer encoder to analyze both anterior segment HR-OCT scans and slit lamp photos, as well as a CNN to perform segmentation of the ocular surface epithelium to establish a diagnosis of OSSN (Figure 1). The DL models can collect data from both anterior segment HR-OCT and conventional slit lamp photos to predict the likelihood of OSSN. In addition, a segmentation tool can be used in conjunction with the DL OCT classification tool to establish a diagnosis of OSSN according to the process outlined in Figure 1. The slit lamp photo artificial intelligence (Al) process can use a vision transformer encoder and a classification head. However, it is different than the DL analysis of HR-OCT scans in that before applying the CNN classification tool, an Al can first be applied to assess quality of the slit lamp photo(s)to ensure it meets appropriate standards for analysis.

[0039] Embodiments of the subject invention provide a focused technical solution to the focused technical problem of how to non-invasively diagnose ocular afflictions (e.g., OSSN). The solution is provided by using a DL model to analyze an ocular image (e.g., HR-OCT scan, slit lamp photograph, or other photograph) to determine from the scan whether a patient has an ocular affliction (e.g., OSSN or any other eye affliction) or not. The DL model can be trained (e.g., using a large number (e.g., at least 50,000, at least 60,000, at least 70,000, or at least 80,000) ocular images, such as HR-OCT scans or slit lamp photographs). Embodiments of the subject invention can improve the computer system performing the DL-aided diagnosis by increasing the efficiency and efficacy of the diagnosis, for example by using a well-trained DL model such that diagnosis proceeds more quickly (which can free up memory and / or processor usage).

[0040] The methods and processes described herein can be embodied as code and / or data. The software code and data described herein can be stored on one or more machine-readable media (e.g., computer-readable media), which may include any device or medium that can store code and / or data for use by a computer system. When a computer system and / or processor reads and executes the code and / or data stored on a computer-readable medium, the computer system and / or processor performs the methods and processes embodied as data structures and code stored within the computer-readable storage medium.

[0041] It should be appreciated by those skilled in the art that computer-readable media include removable and non-removable structures / devices that can be used for storage of information, such as computer-readable instructions, data structures, program modules, and other data used by a computing system / envirorrment. A computer-readable medium includes, but is not limited to, volatile memory such as random access memories (RAM, DRAM, SRAM): and non-volatile memory such as flash memory, various read-only-memories (ROM, PROM, EPROM, EEPROM), magnetic and ferromagnetic / ferroelectric memories (MRAM, FeRAM), and magnetic and optical storage devices (hard drives, magnetic tape, CDs, DVDs); network devices; or other media now known or later developed that are capable of storing computer-readable information / data. Computer-readable media should not be construed or interpreted to include any propagating signals. A computer-readable medium of embodiments of the subject invention can be, for example, a compact disc (CD), digital video disc (DVD), flash memory device, volatile memory, or a hard disk drive (HDD), such as an external HDD or the HDD of a computing device, though embodiments are not limited thereto. A computing device can be, for example, a laptop computer, desktop computer, server, cell phone, or tablet, though embodiments are not limited thereto.

[0042] When ranges are used herein, combinations and subcombinations of ranges (including any value or subrange contained therein) are intended to be explicitly included. When the term “about” is used herein, in conjunction with a numerical value, it is understood that the value can be in a range of 95% of the value to 105% of the value, i.e. the value can be + / - 5% of the stated value. For example, “about 1 kg” means from 0.95 kg to 1.05 kg.

[0043] A greater understanding of the embodiments of the subject invention and of their many advantages may be had from the following examples, given by way of illustration. The following examples are illustrative of some of the methods, applications, embodiments, and variants of the present invention. They are, of course, not to be considered as limiting the invention. Numerous changes and modifications can be made with respect to embodiments of the invention.

[0044] MATERIALS AND METHODS

[0045] Data collection

[0046] A retrospective study was performed with patients who received ocular anterior segment HR-OCT (Optovue, Fremont, CA) scans in the ocular surface oncology clinic at Bascom Palmer Eye Institute (Miami, FL). The study received approval from the University of Miami Institutional Review Board and all research adhered to the tenets of the Declaration of Helsinki.

[0047] All raw anterior OCT imaging data were extracted from the dedicated HR-OCT device in the ocular surface oncology clinic, which to the best of our knowledge represents the largest known dataset of HR-OCT scans for ocular surface tumors. The data was matched to each patient’s medical record number and demographics and clinical diagnoses were retrieved from the electronic health record (EHR) and manually classified under the best-suited category: “Normal”; “OSSN”; “Pterygium”; “Pinguecula”; “Multiple”; “Unknown”; or “Other”. If the categories “Multiple”, “Unknown”, or “Other” were assigned, additional information was collected about the established, or most likely, diagnosis.

[0048] Patients were included if they had one of the following clinical or biopsy-proven diagnoses: OSSN; pinguecula; pterygium; primary acquired melanosis (PAM); conjunctival cyst; conjunctival nevus; lymphoma; and conjunctival melanoma. Exclusion criteria were patient age less than 18 years old, infectious / inflammatory diagnoses, history of corneal transplants, history of keratoconus, and lesions classified as “Unknown”, for which no clear diagnosis was established.

[0049] HR-OCT scans

[0050] Each HR-OCT scan was comprised of multiple cross-sectional images of the ocular surface, also known as B-scans, along identifying lines on corresponding infrared images of the anterior surface (see also Figure 2B), where each B-scan was obtained. Multiple B-scans were manually obtained and aggregated during every HR-OCT scan to assess the entire lesion. Each HR-OCT scan was extracted from the database as raw data files (.OCT files) and de-identified, converted to image files (.PNG), and stored in a highly protected HIPAA-compliant virtual network space.

[0051] Labeling process

[0052] The cleaned dataset included randomly selected HR-OCT scans from the sample for use in data labeling. Only scans from patients’ initial visits were included to avoid training the DL with lesions that had undergone medical treatment or biopsy. For each scan, half of the available B-scans were included to reduce the quantity of scans that were labeled while maintaining the same number of cases. This dataset included the same proportion of diagnoses from the original patient sample to establish a representative pool of eyes typically scanned using HR-OCT at the institution. Finally, a test dataset was created using a random sample of HR-OCT scans with histopathological evidence of the disease to examine diagnostic sensitivity and specificity. All scans were divided at the patient level for subsequent analysis.

[0053] A labeling tool was built using Python programming language for graders to assign a diagnosis to each HR-OCT B-scan for DL algorithm training. The labeling tool included features such as a bounding box to localize the lesion in each B-scan, the corresponding infrared image and line for each B-scan was taken on the ocular surface, and optional image filters (manual settings such as adaptive contrast, brightness, and super resolution filters based on generative adversarial networks) for the graders to use for enhanced visualization of the HR-OCT image. The tool allowed for multiple diagnoses to be assigned to each HR-OCT B-scan amongst the following categories: “Normal”; “OSSN”; “Suspicious for OSSN”; “Pterygium”; “Pinguecula”; “Nevus”; “PAM”; or “Unknown”. Clinical diagnoses extracted from the EHR were provided to assist in the accuracy in diagnosis assignment and obtain the best possible clinical label for training.

[0054] Three cornea specialists (referred to as “GA”, “AN”, and “SP”, respectively) experienced in reading HR-OCT images were identified to label each scan and assign diagnoses for each B- scan using the labeling tool as described above. Each HR-OCT scan was seen by two graders to establish reproducibility in the model and individual B-scan with agreed-upon diagnoses were kept for model training. The “Unknown” diagnosis labels and B-scan that were flagged as low- quality scans were discarded. B-scans that did not receive the same diagnostic labels between the two graders were reviewed again by two senior cornea specialists with vast experience reading HR-OCT images (referred to as “CK” and “AG”, respectively). If scans were still classified as unknown after the additional review, they were discarded.

[0055] DL classification model

[0056] The DL model was constructed in a two-step procedure. First, unlabeled data was used to train a masked autoencoder to reduce the image into a low-dimensional latent space. Second, the randomly selected images for labeling were used as input for fine-tuning of the binary classifier.

[0057] All of OCT images extracted from the machine were used to train the autoencoder for use as an input in the supervised learning model, with the exception of images used in the training, validation, and testing groups of the vision transformer (ViT) model in the second step of the supervised training. The autoencoder, employing an encoder-decoder architecture, is rooted in a fundamental model (see Zhou et al., A foundation model for generalizable disease detection from retinal images, Nature, 2023, 622(7981): p. 156-163; which is hereby incorporated by reference herein in its entirety), and its configuration followed certain specifications (see He et al., Masked Autoencoders Are Scalable Vision Learners, in Computer Science, 2021, Cornell: Arxiv; which is hereby incorporated by reference herein in its entirety). The encoding process utilized a ViT- base, characterized by 24 Transformer blocks and an embedding vector size of 1,024. The decoder also adopted the ViT called ViT-small, with eight transformer blocks, and an embedding vector size of 512. The encoder took unmapped patches, each with a size of 16 x 16, as inputs and transformed them into a feature vector of size 1,024. The 24 Transformer blocks, incorporating multiheaded self-attention and multi-layer perceptron, processed these feature vectors to generate high-level features. For the decoder masked dummy patches, a ratio of 0.85 was used. The same setting described in the foundational model was used for training, which consisted of 800 epochs, where the first 10 were used for learning rate warm-up (from 0 to 0.001). Checkpoints saving was done after every epoch and used afterwards to fine tune this specific model. This autoencoder was trained using 2 RTX A6000 GPUs (NVIDIA, Santa Clara, CA), each one with 48 gigabytes (GB) of VRAM.

[0058] Next, a ViT supervised model used labeled data as an input for fine-tuning a binary classifier (OSSN versus non-OSSN lesions). The labeled data sample was divided at the patient level into a training dataset, a validation dataset, and a final test dataset. The training dataset included both clinical and biopsy-proven diagnoses to generate the largest possible sample for model trainings. However, the test dataset included only biopsy-proven lesions. The inter-grader agreement was calculated using a kappa score, assessing the percentage of agreed upon diagnoses that were completed by the graders, at the patient-level. Data augmentation such as horizontal inversion and tilting of images was performed to increase heterogeneity of the scans, reducing the possibility of overfitting and allowing the algorithm to learn the most relevant features. The best performing model was selected according to accuracy on the validation dataset and used for all subsequent analysis.

[0059] Data Analysis

[0060] The results of the model were analyzed at the scan-level using a separate test set with 566 scans of lesions with biopsy-proven diagnoses of OSSN and pterygium / pinguecula. Importantly, although the training and validation datasets included both clinical-only and biopsy-proven diagnoses, the test dataset included only biopsy-proven lesions. The presence of an OSSN lesion in each individual scan was verified by the expert clinicians (CK, AG), since the imaging area could extend beyond the lesion. The graders were blinded to the EHR diagnoses on the test set, and the inter-grader agreement was calculated using a kappa score, assessing the percentage of agreed upon diagnoses of patients.

[0061] The diagnostic precision of the DL algorithm was investigated using an Area Under the Receiver Operating Characteristic (ROC) curve (AUC) for the testing dataset. The AUC served as a metric to gauge the diagnostic accuracy of each parameter, with a value of 1.0 denoting perfect discrimination and 0.5 indicating chance discrimination. Given the presence of multiple images from the same eye and, in some instances, both eyes of the same patient in the dataset, a bootstrap resampling procedure was implemented. This resampling considered the patient level as the unit of resampling to address the correlation between multiple measurements within the same patient and provided 95% confidence intervals (CI) and P-values (see also Medeiros et al., A statistical approach to the evaluation of covariate effects on the receiver operating characteristic curves of diagnostic tests in glaucoma, Invest Ophthalmol Vis Sci, 2006, 47(6): p. 2520-7; which is hereby incorporated by reference herein in its entirety). The diagnostic sensitivity of each method (labelers vs. DL model) was further evaluated after matching their specificities using the McNemar test. EXAMPLE 1

[0062] Data extracted from the HR-OCT machine included scans from 4,601 patients comprised of 110,973 images. After applying the inclusion and exclusion criteria, 1771 patients remained in the final sample. Based on clinical notes, 22.5% had a diagnosis of OSSN, 26.8% pterygium, 6.4% pinguecula, 22.6% multiple diagnoses, and 21.6% other conjunctival lesions.

[0063] A total of 2,921 HR-OCT scans (544 patients) were separated from the dataset at the patient level for labeling. The remaining images (105,860 scans) were used in the unsupervised autoencoder model development, achieving a final training loss of 0.192 (validation loss=0.192). The labeled dataset was reviewed by the graders, with 32.6% (N=858) identified with OSSN and 33.1% (N=871) with pterygium / pinguecula. 47.4% (N=899) had other conjunctival lesions and were excluded from the labeled dataset. Of note, the inter-grader agreement was of 72% (kappa=0.15) for the diagnosis of OSSN at the patient level. A subset of 1,456 HR-OCT labeled images of 379 patients (461 eyes) were then used to fine-tune the autoencoder predictions, using the supervised ViT model. The training loss of the supervised model was 0.319 (validation loss=0.371).

[0064] The DL model and grader’s performance was further compared on a subset of 540 HR- OCT scans of 48 patients, not used in the model training, where all patients had a biopsy -proven diagnosis of OSSN (27 subjects, 56%) or pterygium / pinguecula (21 subjects, 44%). The characteristics of the testing sample are outlined in Figure 6. The performance of the DL model and the graders are summarized in Figure 7. The best performing DL model had an AUC of 94.5% (95%CI: 91.8-97.2%), compared with 68.8% (95%CI: 56.2-81.3%) for the graders (P0.001). The DL model had a higher diagnostic accuracy (90.3% vs. 86.2%) and sensitivity (86.4% vs. 69.8%), and lower specificity (93.2% vs. 98.5%) for identifying an OSSN lesion in a HR-OCT scan, compared with the graders, respectively. When the specificity of the DL method was matched at the grader’s specificity (98.5%), the DL model had a slightly higher, but non- significantly different sensitivity of 72.3% (95%CI: 66.2%-77.9%) from the graders (P=0.486; McNemar Test).

[0065] A model was also developed for evaluation of slit lamp images instead of the HR-OCT images. 52,533 images were first put through a quality selection, and divided by biopsy proven diagnosis. A resultant 3,772 images were used to fine tune the model. When compared to images given to general eye care providers, the Al model had higher accuracy. Figure 8 shows the results of using the DL model with slit lamp photographs instead of HR-OCT scans. 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.

[0066] All patents, patent applications, provisional applications, and publications referred to or cited herein (including in the “References” section, if present) 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.

[0067] REFERENCES Sayed-Ahmed, 1.0. , et al., Diagnosis and Medical Management of Ocular Surface Squamous Neoplasia. Expert Rev Ophthalmol, 2017. 12(1): p. 11-19. Shields, C.L., et al., Conjunctival Tumors: Review of Clinical Features, Risks, Biomarkers, and Outcomes— The 2017 J. Donald M. Gass Lecture. Asia Pac J Ophthalmol (Phila), 2017. 6(2): p. 109-120. McClellan, A. J., et al., Epidemiology of Ocular Surface Squamous Neoplasia in a Veterans Affairs Population. Cornea, 2013. 32(10): p. 1354-8. Gichuhi, S., et al., Epidemiology of ocular surface squamous neoplasia in Africa. Trop Med Int Health, 2013. 18(12): p. 1424-43. Patel, U., C.L. Karp, and S.R. Dubovy, Update on the Management of Ocular Surface Squamous Neoplasia. Curr Ophthalmol Rep, 2021. 9(1): p. 7-15. Huang, D., et al., Optical coherence tomography. Science, 1991. 254(5035): p. 1178- 81. Izatt, J. A., et al., Micrometer-scale resolution imaging of the anterior eye in vivo with optical coherence tomography. Arch Ophthalmol, 1994. 112(12): p. 1584-9. Rodriguez, M., et al., High-Resolution Optical Coherence Tomography in the Differentiation of Inflammatory Versus Noninflammatory Peripheral Corneal Thinning. Cornea, 2017. 36(1): p. 48-52. Venkateswaran, N., et al., High resolution anterior segment optical coherence tomography of ocular surface lesions: A review and handbook. Expert Rev Ophthalmol, 2021. 16(2): p. 81-95. Vajzovic, L.M., et al., Ultra high-resolution anterior segment optical coherence tomography in the evaluation of anterior corneal dystrophies and degenerations. Ophthalmology, 2011. 118(7): p. 1291-6. Nanji, A.A., et al., High-Resolution Optical Coherence Tomography as an Adjunctive Tool in the Diagnosis of Corneal and Conjunctival Pathology. Ocul Surf, 2015. 13(3): p. 226-35. Tran, A.Q., et al., Utility of high-resolution anterior segment optical coherence tomography in the diagnosis and management of sub-clinical ocular surface squamous neoplasia. Eye Vis (Lond), 2019. 6: p. 27. Karp, C.L., et al., Use of High-Resolution Optical Coherence Tomography in the Surgical Management of Ocular Surface Squamous Neoplasia: A Pilot Study. Am J Ophthalmol, 2019. 206: p. 17-31. 14. Xu, Y., et al., The clinical value of in vivo confocal microscopy for diagnosis of ocular surface squamous neoplasia. Eye (Lond), 2012. 26(6): p. 781-7.

[0068] 15. Hollhumer, R., S. Williams, and P. Michelow, Ocular surface squamous neoplasia: management and outcomes. Eye (Lond), 2021. 35(6): p. 1562-1573.

[0069] 16. Kozma, K., et al., Topical pharmacotherapy for ocular surface squamous neoplasia: systematic review and meta-analysis. Sci Rep, 2022. 12(1): p. 14221.

[0070] 17. Wu, X., et al., Application of artificial intelligence in anterior segment ophthalmic diseases: diversity and standardization. Ann Transl Med, 2020. 8(11): p. 714.

[0071] 18. Medeiros, F.A., A.A. Jammal, and A.C. Thompson, From Machine to Machine: An OCT-Trained Deep Learning Algorithm for Objective Quantification of Glaucomatous Damage in Fundus Photographs. Ophthalmology, 2019. 126(4): p. 513-521.

[0072] 19. Gu, H., et al., Deep learning for identifying corneal diseases from ocular surface slitlamp photographs. Sci Rep, 2020. 10(1): p. 17851.

[0073] 20. Gheisari, S., et al., A combined convolutional and recurrent neural network for enhanced glaucoma detection. Scientific Reports, 2021. 11(1): p. 1945. zi . Noimando, E.ivl., et al., A CNN-aided method to predict glaucoma progression using DARC (Detection of Apoptosing Retinal Cells). Expert Rev Mol Diagn, 2020. 20(7): p. 737-748.

[0074] 22. Tiwari, M., et al., Differentiation of Active Corneal Infections from Healed Scars Using Deep Learning. Ophthalmology, 2022. 129(2): p. 139-146.

[0075] 23. Redd, T.K., et al., Image-Based Differentiation of Bacterial and Fungal Keratitis Using Deep Convolutional Neural Networks. Ophthalmol Sci, 2022. 2(2): p. 100119.

[0076] 24. Hung, N., et al., Using Slit-Lamp Images for Deep Learning-Based Identification of Bacterial and Fungal Keratitis: Model Development and Validation with Different Convolutional Neural Networks. Diagnostics (Basel), 2021. 11(7).

[0077] 25. Zhou, Y., et al., A foundation model for generalizable disease detection from retinal images. Nature, 2023. 622(7981): p. 156-163.

[0078] 26. Kaiming He, X.C., Saining Xie, Yanghao Li, Piotr Dollar, Ross Girshick, Masked Autoencoders Are Scalable Vision Learners, in Computer Science. 2021, Cornell: Arxiv.

[0079] 27. Medeiros, F.A., et al., A statistical approach to the evaluation of covariate effects on the receiver operating characteristic curves of diagnostic tests in glaucoma. Invest Ophthalmol Vis Sci, 2006. 47(6): p. 2520-7.

[0080] 28. Abrams, L.S., et al., Agreement among optometrists, ophthalmologists, and residents in evaluating the optic disc for glaucoma. Ophthalmology, 1994. 101(10): p. 1662-7.

[0081] J:\UM\108XClPCT\Application\Application-asfiled.docx 29. Atallah, M., et al., Role of high resolution optical coherence tomography in diagnosing ocular surface squamous neoplasia with coexisting ocular surface diseases. Ocul Surf, 2017. 15(4): p. 688-695.

[0082] 30. Jampel, H.D., et al., Agreement among glaucoma specialists in assessing progressive disc changes from photographs in open-angle glaucoma patients. Am J Ophthalmol, 2009. 147(1): p. 39-44 el .

[0083] 31. Varma, R., W.C. Steinmann, and I.U. Scott, Expert agreement in evaluating the optic disc for glaucoma. Ophthalmology, 1992. 99(2): p. 215-21.

[0084] 32. Al-Aswad, L.A., et al., Artificial Intelligence for Glaucoma: Creating and Implementing Artificial Intelligence for Disease Detection and Progression. Ophthalmol Glaucoma, 2022. 5(5): p. el6-e25.

[0085] 33. Heo, T.Y., et al., Development of a Deep-Learning-Based Artificial Intelligence Tool for Differential Diagnosis between Dry and Neovascular Age-Related Macular Degeneration. Diagnostics (Basel), 2020. 10(5).

[0086] 34. Diaz-Pinto, A., et al., CNNsfor automatic glaucoma assessment using fundus images: an extensive validation. BioMedical Engineering OnLine, 2019. 18(1): p. 29.

[0087] 35. Kim, K.M., et al., Development of a Fundus Image-Based Deep Learning Diagnostic Tool for Various Retinal Diseases. J Pers Med, 2021. 11(5).

[0088] 36. Channa, R., et al., Effectiveness of artificial intelligence screening in preventing vision loss from diabetes: a policy model. NPJ Digit Med, 2023. 6(1): p. 53.

[0089] 37. Yousefi, S., Clinical Applications of Artificial Intelligence in Glaucoma. J Ophthalmic Vis Res, 2023. 18(1): p. 97-112.

[0090] 38. Yoo, T.K., et al., Adopting low-shot deep learning for the detection of conjunctival melanoma using ocular surface images. Computer Methods and Programs in Biomedicine, 2021. 205: p. 106086.

[0091] 39. Koseoglu, N.D., Z.M. Correa, and T.Y.A. Liu, Artificial intelligence for ocular oncology. Curr Opin Ophthalmol, 2023. 34(5): p. 437-440.

[0092] 40. Zhang, EL, et al., Validation of the Relationship Between Iris Color and Uveal Melanoma Using Artificial Intelligence With Multiple Paths in a Large Chinese Population. Front Cell Dev Biol, 2021. 9: p. 713209.

[0093] 41. Jin, K. and J. Ye, Artificial intelligence and deep learning in ophthalmology: Current status and future perspectives. Advances in Ophthalmology Practice and Research, 2022. 2(3): p. 100078.

[0094] 42. Kumar, S. V. and D. Joshi, Ocular surface squamous neoplasia. Med J Armed Forces India, 2018. 74(3): p. 273-275.

[0095] J:\UM\108XClPCT\Application\Application-asfiled.docx Lei Zhou, H.L., Joseph Bae, Junjun He, Dimitris Samaras, Prateek Prasanna, Self Pretraining with Masked Autoencoders for Medical Image Classification and Segmentation, in Image and Video Processing, arXiv, Editor. 2022, Cornell University. Bowd, C., et al., Individualized Glaucoma Change Detection Using Deep Learning Auto Encoder-Based Regions of Interest. Transl Vis Sei Technol, 2021. 10(8): p. 19. Sewani, H. and R. Kashef, An Autoencoder-Based Deep Learning Classifier for Efficient Diagnosis of Autism. Children (Basel), 2020. 7(10). Raghavendra, U., et al., A Two Layer Sparse Autoencoder for Glaucoma Identification with Fundus Images. J Med Syst, 2019. 43(9): p. 299. Atteia, G., N. Abdel Samee, and H. Zohair Hassan, DFTSA-Net: Deep Feature Transfer-Based Stacked Autoencoder Network for DME Diagnosis. Entropy (Basel), 2021. 23(10). C 1 PCT\Appl ication\Application-asfiled.docx

Claims

CLAIMSWhat is claimed is:

1. A system for deep learning (DL)-based diagnosis of an ocular affliction, the system comprising: a processor; and a machine-readable medium in operable communication with the processor and having instructions stored thereon that, when executed by the processor, perform the following steps: a) receiving first data of an image of an eye of a patient; b) using a convolutional neural network (CNN) to perform segmentation on the first data; c) using the CNN to determine second data of a first characteristic of the eye, the second data comprising at least one of a peak of the first characteristic of the eye and an average of the first characteristic of the eye; and d) comparing the second data to a trained DL model to make a determination of whether the patient has the ocular affliction or not.

2. The system according to claim 1, the segmentation being epithelial segmentation, and the first characteristic being epithelial thickness.

3. The system according to any of claims 1-2, wherein the ocular affliction is ocular surface squamous neoplasia (OSSN).

4. The system according to any of claims 1-3, wherein the image of the eye is a high resolution optical coherence tomography (OCT) (HR-OCT) image.

5. The system according to any of claims 1-3, wherein the image of the eye is a photographic ocular image.

6. The system according to any of claims 1-5, wherein the instructions when executed further comprise, before step d), training a DL model using at least 20,000 ocularimages with known diagnostic information to obtain the trained DL model.

7. The system according to claim 6, wherein the training of the DL model comprises using at least 80,000 ocular images with known diagnostic information.

8. The system according to any of claims 6-7, wherein the ocular images used for training of the DL model are HR-OCT images.

9. The system according to any of claims 6-7, wherein the ocular images used for training of the DL model are photographic ocular images.

10. The system according to any of claims 6-9, wherein the known diagnostic information comprises annotations from expert graders.

11. The system according to any of claims 6-10, wherein the training of the DL model comprises using an encoder and a custom head classifier.

12. The system according to claim 11 , wherein the custom head classifier comprises a vision transformer supervised model.

13. The system according to any of claims 1-12, wherein the second data comprises the peak of the first characteristic of the eye and the average of the first characteristic of the eye.

14. The system according to any of claims 1-13, wherein the determination of whether the patient has the ocular affliction or not is based on if the second data is within a predetermined threshold from a benchmark value from the trained DL model.

15. The system according to any of claims 1-14, wherein the determination of whether the patient has the ocular affliction or not comprises using a vision transformer.

16. The system according to any of claims 1-15, further comprising a display in operable communication with at least one of the processor and the machine-readable medium,J:\UM\108XClPCT\Application\Application-asfiled.docxwherein the instructions when executed further comprise: e) displaying, on the display, the determination of whether the patient has the ocular affliction or not.

17. A method for deep learning (DL)-based diagnosis of an ocular affliction, the method comprising: a) receiving first data of an image of an eye of a patient; b) using a convolutional neural network (CNN) to perform segmentation on the first data; c) using the CNN to determine second data of a first characteristic of the eye, the second data comprising at least one of a peak of the first characteristic of the eye and an average of the first characteristic of the eye; and d) comparing the second data to a trained DL model to make a determination of whether the patient has OSSN or not.

18. The method according to claim 17, the segmentation being epithelial segmentation, and the first characteristic being epithelial thickness.

19. The method according to claim 17, wherein the ocular affliction is ocular surface squamous neoplasia (OSSN).

20. The method according to any of claims 17-19, wherein the image of the eye is a high resolution optical coherence tomography (OCT) (HR-OCT) image.

21. The method according to any of claims 17-19, wherein the image of the eye is a photographic ocular image.

22. The method according to any of claims 17-21, further comprising, before step d), training a DL model using at least 20,000 ocular images with known diagnostic information to obtain the trained DL model.

23. The method according to claim 22, wherein the training of the DL model comprises using at least 80,000 ocular images with known diagnostic information.

24. The method according to any of claims 22-23, wherein the ocular images used for training of the DL model are HR-OCT images.

25. The method according to any of claims 22-23, wherein the ocular images used for training of the DL model are photographic ocular images.

26. The method according to any of claims 22-25, wherein the known diagnostic information comprises annotations from expert graders.

27. The method according to any of claims 22-26, wherein the training of the DL model comprises using an encoder and a custom head classifier.

28. The method according to claim 27, wherein the custom head classifier comprises a vision transformer supervised model.

29. The method according to any of claims 17-28, wherein the second data comprises the peak of the first characteristic of the eye and the average of the first characteristic of the eye.

30. The method according to any of claims 17-29, wherein the determination of whether the patient has the ocular affliction or not is based on if the second data is within a predetermined threshold from a benchmark value from the trained DL model.

31. The method according to any of claims 17-30, wherein the determination of whether the patient has the ocular affliction or not comprises using a vision transformer.

32. The method according to any of claims 17-31, further comprising: e) displaying the determination of whether the patient has the ocular affliction or not.

Images