Detecting cardiovascular disease based on an eye image of a subject

Abstract

Approaches for detecting the presence of coronary artery disease (CAD) using an input eye image of a subject are described. The CAD detection is performed by analyzing the image through machine learning models that detect the presence of at least one of coronary artery stenosis and coronary artery calcium (CAC). Once obtained, the input eye image is processed by two distinct models, i.e., one for stenosis detection and another for CAC detection. Each model extracts values corresponding to a plurality of characteristic attributes from the image, which indicate vascular and morphological details of the fundus of the subject's eye. Based on these extracted values, the models determine the presence of stenosis and / or CAC, which are then used to assess the likelihood or presence of CAD.

Description

DETECTING CARDIOVASCULAR DISEASE BASED ON AN EYE IMAGE OF A SUBJECTBACKGROUND

[0001] Coronary artery disease (CAD) is primarily indicated by two key factors, i.e. , stenosis, which refers to the narrowing of blood vessels, and the presence of calcified plaque, measure through coronary artery calcium (CAC) scoring. Stenosis may significantly restrict blood flow, while calcium deposits in the arterial walls reflect the extent of atherosclerotic burden. Together, these markers provide critical insights into the progression and severity of CAD, a condition where reduced blood supply to the heart increases the risk of heart attacks and other cardiovascular events. CAD is a leading cause of morbidity and mortality globally, often progressing silently until severe complications arise. Early detection of CAD is critical for timely intervention and improved outcomes. Conventionally, CAD and its indicators are diagnosed using invasive methods like coronary angiography or non-invasive techniques such as CT angiograms, stress tests, and CAC scoring, which quantifies calcified plaque to aid in risk stratification and preventive care.BRIEF DESCRIPTION OF FIGURES

[0002] Systems and / or methods, in accordance with examples of the present subject matter are now described and with reference to the accompanying figures, in which:

[0003] FIG. 1 illustrates a training system for training a stenosis detection model and a coronary artery calcium (CAC) detection model, as per an example;

[0004] FIG. 2 illustrates a detection system for detecting presence of coronary artery disease within a subject using at least one of a trained stenosis detection model and a trained CAC detection model, as per an example;

[0005] FIG. 3 illustrates a method for training a stenosis detection model and a CAC detection model, as per an example;

[0006] FIGS. 4A-4B illustrates a method for detecting presence of coronary artery disease within a subject using at least one of a trained stenosis detection model and a trained CAC detection model, as per an example; and

[0007] FIG. 5 illustrates a system environment implementing a non- transitory computer readable medium for detecting presence of coronary artery disease within a subject using at least one of a trained stenosis detection model and a trained CAC detection model, as per an example.DETAILED DESCRIPTION

[0008] Coronary artery disease (CAD) is a leading cause of death and disability worldwide. CAD arises from progressive narrowing or blockage of coronary arteries, which supply oxygen-rich blood to the heart muscle. Two central pathological features indicative of CAD are stenosis, defined as the abnormal constriction of blood vessels due to the accumulation of atherosclerotic plaques, and the presence of calcified plaque, which can be quantified using coronary artery calcium (CAC) scoring. Stenosis impairs myocardial perfusion, while calcium deposits reflect the overall burden of atherosclerosis. These conditions may lead to clinical manifestations such as chest pain (angina), myocardial infarction (heart attack), and sudden cardiac death.

[0009] Early detection and accurate assessment of coronary stenosis and calcified plaque burden are considered critical for initiating timely therapeutic interventions, guiding treatment decisions, and preventing lifethreatening cardiovascular events. However, widespread screening of stenosis and calcium plaque burden remains a challenge due to limitations in current diagnostic modalities.

[0010] Conventional diagnostic techniques for detecting coronary stenosis include both invasive and non-invasive procedures. Invasiveprocedure, such as coronary angiography, involves catheter insertion and contrast injection to visualize coronary vessels. While highly accurate, these procedures carry risks such as bleeding, infection, and radiation exposure, and are unsuitable for routine screening. Non-invasive procedures like computed tomography coronary angiography (CTCA) and cardiac magnetic resonance imaging (MRI) offer detailed anatomical insights but require expensive equipment, specialized personnel, and are often inaccessible in low-resource settings.

[0011] Additional methods, such as stress testing, CT-based calcium scoring, and clinical risk calculators (e.g., the Pooled Cohort Equations or PCE) provide indirect assessments of cardiovascular risk. CAC scoring, in particular, offers a non-invasive means to quantify calcified plaque and stratify risk in asymptomatic individuals. However, these tools have limitations in individual-level prediction and may not generalize well across diverse populations. Moreover, conventional approaches tend to detect disease only after substantial progression, missing opportunities for early intervention. Invasive procedures are impractical for asymptomatic individuals, and non-invasive imaging remains cost-prohibitive for large- scale screening. Risk calculators, while useful, rely on demographic and clinical parameters that may not fully capture vascular pathology.

[0012] Approaches for detecting presence of CAD using an input eye image of a subject are described. As described above as well, for detecting CAD, there are two primary factors which needs to be diagnosed (not at the same time, at least one is required), these factors are detecting presence of stenosis and detecting presence of CAC. Therefore, eventually, the detection result of presence of coronary artery stenosis and coronary artery calcification may then be used for subsequently detecting presence of CAD within the subject. In an example, the input eye image may correspond to the subject eye that is being evaluated for detecting presence of CAD. In an example, the input eye image is a fundus image representing the posterior segment of the eye and including vascular structures such as arteries,veins, the optic disc, and the macula. Such input eye image may be either stored in a data repository or may be captured by an imaging device in realtime scenarios.

[0013] Once obtained, the input eye image may be processed by two distinct machine learning models, e.g., one for detecting presence of coronary artery stenosis and another for detecting presence of CAC, to diagnose the subject as having coronary artery disease. Prior to processing, the input eye image may be undergo various image enhancement steps and segmentation steps (which are described later in the description). During processing, each machine learning model extracts values corresponding to a plurality of characteristic attributes from the input eye image. In an example, the values corresponding to the plurality of characteristic attributes indicates vascular and morphological details of the fundus of the subject’s eye, which are relevant for detecting presence of either stenosis or CAC.

[0014] Based on the extracted values, the respective machine learning models detect the presence of stenosis and / or CAC in the coronary artery of the subject. For stenosis detection, the extracted attributes may include a disc & cup height, fractal dimensions, a vessel density, an average width of vessels, distance tortuosity, squared curvature tortuosity, and tortuosity density. For CAC detection, the other machine learning model may utilize additional or overlapping features that are indicative of calcified plaque burden. These attributes collectively reflects vascular health and may indicate systemic changes associated with coronary artery disease. The outputs from both the stenosis detection model and the CAC detection model are subsequently integrated and analyzed by a unified CAD detection engine, which determines the overall presence or likelihood of CAD in the subject. This combined assessment enhances diagnostic accuracy by leveraging both structural and compositional indicators of vascular pathology.

[0015] While the above examples describe the use of two distinct models, i.e., one for stenosis detection and another for CAC detection, it isimportant to note that either model may be used independently to detect the presence of coronary artery disease (CAD). Depending on the performance and diagnostic value of each model, the present approaches may operate using only the stenosis detection model or only the CAC detection model to assess CAD risk.

[0016] Returning to the present example, the above-mentioned machine learning models, e.g., a stenosis detection model and a CAC detection model, are sophisticated deep learning algorithm, each trained on diverse training dataset comprising training fundus images and corresponding annotations or labels for stenosis and CAC, respectively. Such training process enables the machine learning models to learn the complex relationship between the characteristic attributes of the training fundus images and the presence, location, and seventy of stenotic or calcified conditions.

[0017] The training dataset for both machine learning models may include a wide range of stenotic and calcified conditions, anatomical variations, and image quality parameters to ensure robustness and generalizability across different patient populations and imaging conditions. During training, the machine learning models may utilize techniques such as data augmentation, transfer learning, and multi-scale feature extraction to enhance their ability to detect subtle vascular changes. The machine learning models may be trained based on advanced neural network architectures, such as convolutional neural networks (CNNs), vision transformers, or hybrid architectures, to achieve optimal performance in both stenosis and CAC identification and classification tasks.

[0018] The present subject matter provides a non-invasive, scalable, and clinically impactful solution for the detection of coronary artery disease (CAD) through the analysis of retinal fundus images. CAD detection is achieved by leveraging a unified approach that integrates outputs from two machine learning models, i.e., one for detecting coronary artery stenosis and another for detecting coronary artery calcium (CAC). By utilizing theanatomical and physiological similarities between retinal and coronary vasculature, the present approaches identifies vascular abnormalities indicative of CAD, eliminating the need for invasive procedures such as coronary angiography or costly imaging modalities like CT angiography and cardiac MRI. This approach enables early detection by analyzing morphological and vascular features visible in the retina, which reflect systemic cardiovascular conditions including both stenosis and calcified plaque burden.

[0019] The manner in which the machine learning model is trained and used for detecting presence of stenosis within the coronary artery is explained in detail in conjunction with description of FIGS. 1 -5. While aspects of described systems may be implemented in any number of different electronic devices, environments, and / or implementation, the examples are described in the context of the following example device(s). In another example, the aspects of the present subject matter may also be implemented by a standalone device having executable instructions. It may be noted that drawings of the present subject matter shown here are for illustrative purposes and are not to be construed as limiting the scope of the subject matter claimed.

[0020] FIG. 1 illustrates a training system 102 comprising a processor or memory (not shown), for training a stenosis detection model and a CAC (Coronary Artery Calcification) detection model. In an example, the training system 102 (referred to as system 102) may be communicatively coupled to a repository 104 through a network 106. The repository 104 may further include training data 108. The training data 108 may include data that may be used for training both the stenosis detection model and the CAC detection model. In an example, the training data 108 includes training fundus images and corresponding reference stenosis indicators indicating status of stenosis, as well as reference CAC indicators indicating presence of extent of CAC, in the training fundus images for training the respective detection model.

[0021] In one example, the training data 108 may also include training attribute data corresponding to each training fundus image. Such attribute data may include values corresponding to a plurality of characteristic attributes indicating vascular and morphological details of the fundus of the training fundus image. Examples of such plurality of characteristic attributes include, but are not limited to, a disc & cup height, fractal dimensions, a vessel density, an average width of vessels, distance tortuosity, squared curvature tortuosity, and tortuosity density. These attributes may be relevant for training both the stenosis detection model and the CAC detection model.

[0022] In addition, the training data 108 may also include training health marker data associated with each training fundus image. The training health marker data includes values corresponding to a plurality of health marker associated with each training fundus image. Examples of such plurality of health marker include, but are not limited to, age, gender, height, weight, smoking status, blood pressure value, cholesterol levels, and history of cardiovascular diseases. The above-mentioned attribute data or health marker data is carefully annotated with the corresponding training reference stenosis indicator and / or CAC indicator, indicating status of stenosis and / or CAC within each training fundus image, to provide comprehensive training information for the stenosis detection model and the CAC detection model.

[0023] Although depicted as being obtained from a single repository, such as repository 104, the training data 108 may also be obtained from multiple other sources without deviating from the scope of the present subject matter. In such cases, each of such multiple repositories may be interconnected through a network, such as network 106.

[0024] The network 106 may be a private network or a public network and may be implemented as a wired network, a wireless network, or a combination of a wired and wireless network. The network 106 may also include a collection of individual networks, interconnected with each other and functioning as a single large network, such as the Internet. Examples of such individual networks include, but are not limited to, Global System forMobile Communication (GSM) network, Universal Mobile Telecommunications System (UMTS) network, Personal Communications Service (PCS) network, Time Division Multiple Access (TDMA) network, Code Division Multiple Access (CDMA) network, Next Generation Network (NGN), Public Switched Telephone Network (PSTN), Long Term Evolution (LTE), and Integrated Services Digital Network (ISDN).

[0025] The system 102 may further include instructions 110 and a training engine 112. In an example, the instructions 110 are fetched from memory and executed by a processor included within the system 102. The training engine 112 may be implemented as a combination of hardware and programming, for example, programmable instructions to implement a variety of functionalities. In examples described herein, such combinations of hardware and programming may be implemented in several different ways. For example, the programming for the training engine 112 may be executable instructions, such as instructions 110. Such instructions may be stored on a non-transitory machine-readable storage medium which may be coupled either directly with the system 102 or indirectly (for example, through networked means). In an example, the training engine 112 may include a processing resource, for example, either a single processor or a combination of multiple processors, to execute such instructions. In the present examples, the non-transitory machine-readable storage medium may store instructions, such as instructions 110, that when executed by the processing resource, implement training engine 112. In other examples, the training engine 112 may be implemented as electronic circuitry.

[0026] The instructions 110, when executed by the processing resource, cause the training engine 112 to train both the stenosis detection model 114 and the CAC detection model 116 based on the training data 108. The system 102 may further include a training fundus image(s) 118, a training attribute data 120, a training health marker data 122, a reference stenosis indicator data 124, and a reference CAC indicator data 126. In an example, the system 102 may obtain training data 108 corresponding to a singletraining fundus image from the repository 104, and the data pertaining to that is stored as training fundus image(s) 118, training attribute data 120, training health marker data 122, reference stenosis indicator data 124, and reference CAC indicator data 126.

[0027] As described previously, the stenosis detection model 114 and the CAC detection model 116 are machine learning models, and may be implemented as deep learning models. Although the present examples have been described in relation to deep learning models, the aforementioned approaches may also be implemented using other machine learning models. It may also be noted that any explanation provided in conjunction with deep learning models is applicable to other machine learning models, without limitations and without deviating from the scope of the present subject matter. Such examples have not been described in this description for the sake of brevity.

[0028] In operation, the training engine 112 obtains training data 108 including training fundus image(s) 118 and corresponding reference stenosis indicator data 124 and reference CAC indicator data 126 from the repository 104. In an example, the training fundus image(s) 118 refers to retinal images captured from the posterior segment of the eye, including anatomical regions such as the optic disc, macula, and vascular structures like arteries and veins. These images may be acquired using fundus cameras or smartphone-based retinal imaging devices during routine ophthalmic or cardiovascular screenings. The reference stenosis indicator data 124 represent clinically validated information indicating the presence, absence, or seventy of coronary artery stenosis, while the reference CAC indicator data 126 represents clinically validated information indicating the presence or extent or location of coronary artery calcification (CAC). These reference data may be derived from invasive procedures such as coronary angiography or non-invasive imaging techniques like CT coronary angiography, and may include quantitative measures such as percentagenarrowing, calcium scores, or categorical assessments (e.g., mild, moderate, severe).

[0029] The training fundus image(s) 118 are obtained from hospital databases, clinical trials, or screening programs where retinal imaging is performed alongside cardiovascular evaluations. The corresponding reference stenosis indicator data 124 and the reference CAC indicator data 126 are sourced from the same subjects, ensuring alignment between retinal features and both coronary artery stenosis status and CAC status. In one example, the training fundus image(s) 118 may be organized in paired form, with each image in the pair corresponding to either the left or right eye of the subject. In such cases, each image pair is annotated with a single reference stenosis indicator and a reference CAC indicator representing the subject’s overall coronary artery stenosis and CAC condition. Examples of such indicator include, but are not limited to, percentage narrowing, calcium score, location, or any other information relevant to stenosis or CAC.

[0030] In an example, as described above as well, the obtained training data 108 may also include training attribute data 120 and training health marker data 122. In another example, the training attribute data 120 may not be part of the training data 108. In such a case, the training engine 112 processes the training fundus image(s) 118 to extract training attribute data 120 including values corresponding to the plurality of characteristics attribute for each training fundus image of the training fundus image(s) 118. These attributes may be used for training both the stenosis detection model and the CAC detection model.

[0031] Returning to the present example, once obtained, the training engine 112 performs a plurality of image enhancement steps on the training fundus image(s) 118. In an example, the plurality of image enhancement steps is performed to improve the visual quality and diagnostic relevance of the retinal features captured in the training fundus image(s) 118. These enhancements help in highlighting subtle vascular and anatomical details that may be critical for accurate feature extraction and model training.Examples of such image enhancement steps include, but are not limited to, contrast enhancement to improve vessel visibility, sharpness adjustment to clarify structural boundaries, brightness adjustment to normalize illumination, green channel extraction to emphasize vascular structures, and bilateral filtering to reduce noise while preserving edges. These preprocessing steps ensure that the training images are standardized and optimized for downstream analysis, thereby improving the robustness and accuracy of the stenosis detection model 114 and the CAC detection model 116.

[0032] Thereafter, the training engine 112 performs a segmentation process on the training fundus image(s) 118 to extract image segments corresponding to vascular structures and anatomical regions of the eye. In an example, the segmentation process is performed to isolate and delineate specific retinal components that are relevant for assessing cardiovascular health. This includes identifying and separating arteries and veins, which may exhibit morphological changes associated with systemic vascular conditions, as well as anatomical landmarks such as the optic disc and optic cup, which serve as reference points for measuring vessel geometry and distribution. The segmented regions enable targeted feature extraction and facilitate the generation of structured input data for training both the stenosis detection model and the CAC detection model. Accurate segmentation also supports the calculation of derived metrics such as vessel density, tortuosity, and disc-to-cup ratios, which are used to correlate retinal morphology with coronary artery stenosis and CAC.

[0033] It is important to note that the image enhancement and segmentation steps described above are optional and may not be necessary in all cases, depending on factors such as image quality and diagnostic requirements. Accordingly, the present subject matter should not be construed as being limited to the inclusion of these steps. The detection process can be effectively performed with or without them, without departing from the scope of the invention.

[0034] Once image enhancement steps and segmentation process are performed, the training engine 112 may train the stenosis detection model 114 and the CAC detection model 116 based on the training data 108. In an example, the training process involves feeding the enhanced and segmented training fundus image(s) 118, along with the corresponding training attribute data 120, training health marker data 122, reference stenosis indicator data 124, and reference CAC indicator data 126, into the respective models. The stenosis detection model 114 learns to associate specific retinal features and health markers with the presence or severity of coronary artery stenosis, while the CAC detection model 116 learns to associate these features with the presence or extent or location of CAC, as indicated by the respective indicator. During training, each model may optimize its internal parameters to minimize prediction error against the corresponding reference indicators. It may be noted that, the training may be performed over multiple iterations and may include validation steps to assess model performance and prevent overfitting.

[0035] It may be noted that although the above training process describes model development using annotated fundus images and health markers, the present subject matter may also be implemented using pretrained models that are fine-tuned on domain-specific data related to CAD. This approach allows leveraging existing architectures while adapting them to the unique characteristics of retinal imaging for coronary artery disease detection, thereby enhancing model performance and generalizability.

[0036] Once trained, the stenosis detection model 114 is capable of analyzing new, unseen fundus images to detect the presence of coronary artery stenosis in a subject, and the CAC detection model 116 is capable of detecting the presence or extent of CAC. In operation, each model receives an input eye image and processes it to extract relevant vascular and morphological features, such as vessel density, tortuosity, disc and cup dimensions, and other structural indicators. These extracted features are then evaluated in the context of patterns learned during training, allowingthe models to determine whether the subject exhibits signs of narrowing or blockage in one or more coronary arteries (stenosis) and / or the presence of coronary artery calcification.

[0037] In some examples, the models may also incorporate additional inputs such as demographic health markers (e.g., age, gender, blood pressure, cholesterol levels). The output may be presented in various formats, including a binary classification (presence or absence of stenosis or CAC), a severity category (e.g., mild, moderate, severe), or a continuous score reflecting the degree of vascular narrowing or calcification. This enables clinicians to perform non-invasive cardiovascular risk assessments using retinal imaging data, supporting early diagnosis and timely intervention.

[0038] The manner in which the stenosis detection model 114 and the CAC detection model 116 may be used for detecting presence of stenosis and CAC within an input eye image is further described in detailed manner in conjunction with FIG. 2.

[0039] FIG. 2 illustrates an environment 200 with coronary artery disease (CAD) detection system 202 configured to detect presence of CAD by detecting present of at least one of coronary artery stenosis and coronary artery calcification (CAC) within an input eye image. In an example, the CAD detection system 202 (referred to as detection system 202) utilizes a trained stenosis detection model 114 and a CAC detection model 116 to perform respective analysis, which are consequently used for detecting presence of CAD. In an example, the detection system 202 includes a mobile phone, tablet, mac-book, or any other portable computing device having a camera device coupled to it. The detection system 202 obtains the input eye image from a data repository 204 which is connected over a network (not shown in Figure) with the detection system 202. The data repository 204 further includes a plurality of input eye image(s) 206 (referred to as input eye image(s) 206), and a health marker data 208 corresponding to each input eye image.

[0040] In another example, the detection system 202 may obtain input eye images in real-time from a retinal imaging device, such as a smartphone-based fundus camera or a dedicated desktop imaging system. These images may be captured during routine ophthalmic or cardiovascular screenings. Furthermore, the detection system 202 may be integrated directly within the imaging device itself, enabling on-device processing of the captured fundus image. In such an embodiment, the detection system 202 may perform end-to-end analysis, i.e., from image acquisition to stenosis detection and CAC detection, without requiring external computational resources.

[0041] The detection system 202 may include a processor 210, interface(s) 212, and memory(s) 214. The processor 210 may be implemented as microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuitries, and / or other devices that manipulate signals based on operational instructions. Among other capabilities, the processor 210 may be configured to obtain various types of subject’s data, such as eye image, health marker data, and attribute data. The processor 210 may then use a stenosis detection model, such as stenosis detection model 114, to detect presence of stenosis and the CAC detection model 116 to detect presence or extent or location of CAC within the input eye image(s) 206.

[0042] The interface(s) 212 may allow the connection or coupling of the detection system 202 with one or more sensors or devices onboard the detection system 202, depending on the implementation of the detection system 202 through a wired network, a wireless network, or a combination of a wired and wireless network. The interface(s) 212 may also enable intercommunication between different logical as well as hardware components of the detection system 202.

[0043] The memory(s) 214 may be a computer-readable medium, examples of which include volatile memory (e.g., RAM), and / or non-volatile memory (e.g., Erasable Programmable read-only memory, i.e., EPROM,flash memory, etc.). The memory(s) 214 may be an external memory, or internal memory, such as a flash drive, a compact disk drive, an external hard disk driver, or the like. The memory(s) 214 may further include data which either may be utilized or generated during the operation of the detection system 202.

[0044] Similar to the system 102, the detection system 202 may further include instruction(s) 216 and engine(s) 218. In an example, the instruction(s) 216 are fetched from the memory(s) 214 and executed by the processor 210 included within the detection system 202. The engine(s) 218 may include a CAD detection engine 220 and other engine(s) 222. The other engine(s) 222 may further implement functionalities that supplement functions performed by the detection system 202 or any of the engine(s) 218. The CAD detection engine 220 (referred to as detection engine 220) may be implemented as a combination of hardware and programming, for example, programmable instructions to implement a variety of functionalities.

[0045] In examples described herein, such combinations of hardware and programming may be implemented in several different ways. For example, the programming for the detection engine 220 may be executable instructions, such as instruction(s) 216. Such instruction(s) 216 may be stored on a non-transitory machine-readable storage medium which may be coupled either directly with the detection system 202 or indirectly (for example, through networked means). In an example, the detection engine 220 may include a processing resource, for example, either a single processor or a combination of multiple processors, to execute such instructions. In the present examples, the non-transitory machine-readable storage medium may store instructions, such as instruction(s) 216, that when executed by the processing resource, implement the detection engine 220. In other examples, the detection engine 220 may be implemented as electronic circuitry.

[0046] The detection system 202 may further include a stenosis detection model, such as stenosis detection model 114, a CAC detection model 116, and a data 224. The data 224 may include corresponding data that is utilized or generated by the detection system 202, while performing a variety of functions. In an example, the data 224 further includes input eye image(s) 206, health marker data 208, an attribute data 226, stenosis score 228, calcium score 230, CAD detection result 232, and other data 234. Further, the other data 234, amongst other things, may serve as a repository for storing data that is processed, or received, or generated as a result of the execution of the instructions by the processor 210.

[0047] In operation, an input eye image corresponding to the fundus of a subject's eye, intended for analysis to detect the presence of coronary artery stenosis and coronary artery calcification (CAC), is obtained. In an example, the detection engine 220 within the detection system 202 retrieves the input eye image(s) 206 from the data repository 204. The input eye image(s) 206 may be high-resolution fundus photographs captured using a retinal imaging device, such as a fundus camera or smartphone-based retinal imaging system. These images typically depict the posterior segment of the eye, including vascular structures (arteries and veins), the optic disc, and the macula, which are relevant for assessing systemic vascular health.

[0048] It may be noted that, the human retina offers a unique, non- invasive window into systemic vascular health. Retinal blood vessels are the only part of the circulatory system that may be directly visualized without invasive procedures. The retinal vasculature shares embryological origins and physiological characteristics with coronary vessels, making it a promising surrogate for cardiovascular assessment. Changes in retinal vascular morphology, such as vessel caliber, tortuosity, arteriovenous ratio, and branching patterns, may be used as indicators for detecting presence of coronary stenosis. Further, the fundus photography, which captures high- resolution images of the retina, enables detailed analysis of these vascular features. Advanced image processing techniques may be used to extractadditional parameters such as fractal dimensions, vessel density, average vessel width, and various tortuosity metrics. These features may reflect underlying vascular pathology and offer valuable insights into cardiovascular risk.

[0049] Continuing further, once the input eye image(s) 206 is obtained, the detection engine 220 retrieves values corresponding to a plurality of health markers, such as health marker data 208, from the data repository 204. In an example, the health marker data 208 includes subject-specific clinical and demographic information that may influence cardiovascular risk. Examples of such health marker data 208 include, but are not limited to, age, gender, height, weight, smoking status, blood pressure values, cholesterol levels, and history of cardiovascular diseases. These markers are used in conjunction with retinal features to improve the accuracy of stenosis detection.

[0050] Thereafter, the detection engine 220 determines the type of view represented in the input eye image(s) 206. In an example, the type of view indicates the anatomical region captured in the image and may include macula-centered views, optic disc views, or peripheral retinal views. The type of view is determined to ensure that the image contains diagnostically relevant regions necessary for accurate analysis. Based on the determined type of view, the detection engine 220 selectively proceeds to either use the input eye image(s) 206 for further processing or inhibit further processing if the image is deemed unsuitable.

[0051] In cases where the type of view is identified as a maculacentered view or an optic disc view, the detection engine 220 proceeds to use the input eye image(s) 206 for further processing, as these views contain critical vascular and anatomical features. On the other hand, if the image corresponds to other view types that do not provide sufficient diagnostic information, the detection engine 220 inhibits further processing and discards the image. In one example, the detection engine 220 may also generate a prompt on a display device of the detection system 202,indicating that the image is unsuitable for analysis and requesting acquisition of a new image.

[0052] Returning to the present example, once the input eye image(s) 206 is deemed eligible for further processing, the detection engine 220 performs a plurality of image enhancement steps to improve the visibility and clarity of retinal features. These enhancements may include, but are not limited to, contrast enhancement to highlight blood vessels, sharpness adjustment to improve edge definition, brightness normalization to correct illumination inconsistencies, green channel extraction to emphasize vascular structures, and bilateral filtering to reduce noise while preserving important details. These steps ensure that the image is optimized for accurate feature extraction.

[0053] Thereafter, the detection engine 220 performs a segmentation process on the input eye image(s) 206 to extract image segments corresponding to vascular structures and anatomical regions of the subject’s eye. In an example, segmentation is performed to isolate arteries, veins, the optic disc, and the optic cup. These segmented regions are used to compute derived metrics such as vessel density, tortuosity, and disc-to-cup ratios, which are critical indicators of systemic vascular health and potential coronary artery stenosis.

[0054] It is important to note that the image enhancement and segmentation steps described above are optional and may not be necessary in all cases, depending on factors such as image quality and diagnostic requirements. Accordingly, the present subject matter should not be construed as being limited to the inclusion of these steps. The detection process can be effectively performed with or without them, without departing from the scope of the invention.

[0055] Once all relevant data is obtained, derived, and extracted, the detection engine 220 processes the input eye image(s) 206 along with the health marker data 208 and segmented images using the stenosis detection model 114 to detect the presence of stenosis, and using the CAC detectionmodel 116 to detect the presence or extent or location of CAC. In an example, both the models extract values corresponding to a plurality of characteristic attributes, referred to as attribute data 226, from the input eye image(s) 206. The attribute data 226 indicates vascular and morphological details of the fundus of the subject’s eye, including but not limited to disc and cup height and width, vessel density, average vessel width, fractal dimensions, and various tortuosity metrics such as distance tortuosity, squared curvature tortuosity, and tortuosity density. These features are indicative of systemic vascular health and are used to assess the likelihood of narrowing or blockage in the coronary arteries.

[0056] Based on the extracted attribute data 226, the stenosis detection model 114 detects the presence of stenosis, and the CAC detection model 116 detects the presence or extent of CAC in the coronary artery of the subject. In an example, each model may output a binary classification indicating whether the concerned condition is present or not, or a severity level such as mild, moderate, or severe. The extent of stenosis or CAC may be quantified based on the correlation between the extracted retinal features and clinically validated indicators, such as percentage narrowing of the artery or calcium scores. This assessment provides a non-invasive means to evaluate coronary artery health and supports early identification of individuals at risk of cardiovascular events.

[0057] In an example, the stenosis detection model 114 and the CAC detection model 116 may output a binary classification (presence or absence), a categorical seventy level (e.g., mild, moderate, severe), or a continuous score, such as stenosis score 228 and the calcium score 230, indicating the extent of narrowing or calcification. These outputs are determined based on the correlation between retinal features and clinically validated indicators for stenosis and CAC.

[0058] As described above, while the presence of stenosis is a strong indicator of coronary artery disease (CAD), it alone may not confirm the diagnosis. To provide a more comprehensive assessment, the detectionengine 220 obtains both the stenosis score 228, generated by the stenosis detection model 114, and the calcium score 230, generated by the CAC detection model 116 or derived from clinical data, corresponding to the subject from the data repository 204 to detect presence of CAD within the subject. In an example, the stenosis score 228 reflects the likelihood or severity of coronary artery stenosis based on retinal features, while the calcium score 230 indicates the extent of calcified plaque within the coronary arteries, a well-established marker of atherosclerotic burden. As a result of such detection based on stenosis and calcium score (228, 230), the detection engine 220 generates the CAD detection result 232 indicating the overall presence, likelihood, or severity of coronary artery disease in the subject. The CAD detection result 232 may be presented as a binary classification, a categorical risk level (e.g., low, moderate, high), or a continuous score, and may be used to support clinical decision-making and early intervention.

[0059] In one example, these scores, along with other relevant health marker data and attribute data, are provided as input to a dedicated CAD detection model (not shown in FIG. 2), a separate machine learning model trained to integrate multi-modal information. The CAD detection model is configured to analyze the combined outputs from the stenosis detection model 114 and the CAC detection model 116, as well as additional clinical and demographic features, to determine the overall likelihood or presence of CAD within the subject’s cardiovascular system. The CAD detection model may output a binary classification (presence or absence of CAD), a categorical risk level (e.g., low, moderate, high), or a continuous risk score, based on established clinical thresholds and learned correlations.

[0060] This multi-model, multi-modal approach enables detection system 202 to provide a more accurate and robust assessment of CAD risk by leveraging both structural (stenosis) and compositional (calcification) indicators from retinal imaging and clinical data. The combined assessment enhances diagnostic accuracy, supports early identification of individuals atrisk, and facilitates informed clinical decision-making. Additionally, the detection system 202 may generate a comprehensive report summarizing the stenosis score, calcium score, and CAD risk assessment, which may be reviewed by clinicians for further evaluation and management.

[0061] Although the above paragraphs describe the detection of coronary artery disease (CAD) as being based on both stenosis and calcium scores, this should not be considered limiting. The present subject matter may also be implemented using only the stenosis detection model or only the CAC detection model, depending on the diagnostic requirements or model performance. CAD detection using either of these models independently still falls within the scope of the invention and does not deviate from the core principles described herein.

[0062] In another example, the detection system 202 may be communicatively coupled to a central computing server through a network (not shown in FIG. 2). The network may be a private network or a public network and may be implemented as a wired network, a wireless network, or a combination of a wired and wireless network, and may be similar to the network 106 (as depicted in FIG. 1 ). All the above disclosed steps which may be performed by the detection engine 220 of the detection system 202, may be implemented or performed by the central computing server on behalf of the detection system 202 to reduce computing load on edge of the network.

[0063] FIG. 3 illustrates an example method 300 for training a stenosis detection model and a CAC detection model, in accordance with examples of the present subject matter. The order in which the above-mentioned method is described is not intended to be construed as a limitation, and some of the method blocks described may be combined in a different order to implement the method, or alternative method.

[0064] Furthermore, the above-mentioned method may be implemented in a suitable hardware, computer-readable instructions, or combination thereof. The steps of such method may be performed by either a systemunder the instruction of machine executable instructions stored on a non- transitory computer readable medium or by dedicated hardware circuits, microcontrollers, or logic circuits. For example, the method may be performed by a training system, such as system 102. In an implementation, the method may be performed under an “as a service” delivery model, where the system 102, operated by a provider, receives programmable code. Herein, some examples are also intended to cover non-transitory computer readable medium, for example, digital data storage media, which are computer readable and encode computer-executable instructions, where said instructions perform some or all the steps of the above- mentioned method.

[0065] In an example, the method 300 may be implemented by the system 102 for training a stenosis detection model and a CAC detection model based on training data such as training data 108. At block 302, training data including at least training fundus image and a reference stenosis indicator indicating status of stenosis is obtained. For example, the training engine 112 obtains training data 108 including training fundus image(s) 118 and corresponding reference stenosis indicator data 124 and reference CAC indicator data 126 from the repository 104. In an example, the training fundus image(s) 118 refers to retinal images captured from the posterior segment of the eye, including anatomical regions such as the optic disc, macula, and vascular structures like arteries and veins. These images may be acquired using fundus cameras or smartphone-based retinal imaging devices during routine ophthalmic or cardiovascular screenings. The reference stenosis indicator data 124 represent clinically validated information indicating the presence, absence, or seventy of coronary artery stenosis, while the reference CAC indicator data 126 represents clinically validated information indicating the presence or extent or location of coronary artery calcification (CAC). These reference data may be derived from invasive procedures such as coronary angiography or non-invasive imaging techniques like CT coronary angiography, and may includequantitative measures such as percentage narrowing, calcium scores, or categorical assessments (e.g., mild, moderate, severe).

[0066] The training fundus image(s) 118 are obtained from hospital databases, clinical trials, or screening programs where retinal imaging is performed alongside cardiovascular evaluations. The corresponding reference stenosis indicator data 124 and the reference CAC indicator data 126 are sourced from the same subjects, ensuring alignment between retinal features and both coronary artery stenosis status and CAC status. In one example, the training fundus image(s) 118 may be organized in paired form, with each image in the pair corresponding to either the left or right eye of the subject. In such cases, each image pair is annotated with a single reference stenosis indicator and a reference CAC indicator representing the subject’s overall coronary artery stenosis and CAC condition. Examples of such indicator include, but are not limited to, percentage narrowing, calcium score, location, or any other information relevant to stenosis or CAC.

[0067] In an example, as described above as well, the obtained training data 108 may also include training attribute data 120 and training health marker data 122. In another example, the training attribute data 120 may not be part of the training data 108. In such a case, the training engine 112 processes the training fundus image(s) 118 to extract training attribute data 120 including values corresponding to the plurality of characteristics attribute for each training fundus image of the training fundus image(s) 118. These attributes may be used for training both the stenosis detection model and the CAC detection model.

[0068] At block 304, a plurality of image enhancement steps are performed on the training fundus image. For example, the training engine 112 performs a plurality of image enhancement steps on the training fundus image(s) 118. In an example, the plurality of image enhancement steps is performed to improve the visual quality and diagnostic relevance of the retinal features captured in the training fundus image(s) 118. These enhancements help in highlighting subtle vascular and anatomical detailsthat may be critical for accurate feature extraction and model training. Examples of such image enhancement steps include, but are not limited to, contrast enhancement to improve vessel visibility, sharpness adjustment to clarify structural boundaries, brightness adjustment to normalize illumination, green channel extraction to emphasize vascular structures, and bilateral filtering to reduce noise while preserving edges. These preprocessing steps ensure that the training images are standardized and optimized for downstream analysis, thereby improving the robustness and accuracy of the stenosis detection model 114 and the CAC detection model 116.

[0069] At block 306, a segmentation process is performed on the training fundus image to extract image segments corresponding to vascular structures and anatomical regions. For example, the training engine 112 performs a segmentation process on the training fundus image(s) 118 to extract image segments corresponding to vascular structures and anatomical regions of the eye. In an example, the segmentation process is performed to isolate and delineate specific retinal components that are relevant for assessing cardiovascular health. This includes identifying and separating arteries and veins, which may exhibit morphological changes associated with systemic vascular conditions, as well as anatomical landmarks such as the optic disc and optic cup, which serve as reference points for measuring vessel geometry and distribution. The segmented regions enable targeted feature extraction and facilitate the generation of structured input data for training both the stenosis detection model and the CAC detection model. Accurate segmentation also supports the calculation of derived metrics such as vessel density, tortuosity, and disc-to-cup ratios, which are used to correlate retinal morphology with coronary artery stenosis and CAC.

[0070] At block 308, a stenosis detection model and a CAC detection model are trained based on the training data to detect presence of stenosis and calcification, respectively, in a coronary artery of a subject based on aninput eye image of the subject. For example, once image enhancement steps and segmentation process are performed, the training engine 112 may train the stenosis detection model 114 and the CAC detection model 116 based on the training data 108. In an example, the training process involves feeding the enhanced and segmented training fundus image(s) 118, along with the corresponding training attribute data 120, training health marker data 122, reference stenosis indicator data 124, and reference CAC indicator data 126, into the respective models. The stenosis detection model 114 learns to associate specific retinal features and health markers with the presence or seventy of coronary artery stenosis, while the CAC detection model 116 learns to associate these features with the presence or extent or location of CAC, as indicated by the respective indicator. During training, each model may optimize its internal parameters to minimize prediction error against the corresponding reference indicators. It may be noted that, the training may be performed over multiple iterations and may include validation steps to assess model performance and prevent overfitting.

[0071] Once trained, the stenosis detection model 114 is capable of analyzing new, unseen fundus images to detect the presence of coronary artery stenosis in a subject, and the CAC detection model 116 is capable of detecting the presence or extent of CAC. In operation, each model receives an input eye image and processes it to extract relevant vascular and morphological features, such as vessel density, tortuosity, disc and cup dimensions, and other structural indicators. These extracted features are then evaluated in the context of patterns learned during training, allowing the models to determine whether the subject exhibits signs of narrowing or blockage in one or more coronary arteries (stenosis) and / or the presence of coronary artery calcification.

[0072] In some examples, the models may also incorporate additional inputs such as demographic health markers (e.g., age, gender, blood pressure, cholesterol levels). The output may be presented in various formats, including a binary classification (presence or absence of stenosisor CAC), a severity category (e.g., mild, moderate, severe), or a continuous score reflecting the degree of vascular narrowing or calcification. This enables clinicians to perform non-invasive cardiovascular risk assessments using retinal imaging data, supporting early diagnosis and timely intervention.

[0073] FIGS. 4A-4B illustrates example method 400 for detecting presence of stenosis and calcification within a coronary artery based on an input eye image of a subject using a trained stenosis detection model and a trained CAC detection model, respectively. Similar to FIG. 3, the order in which the above-mentioned method is described is not intended to be construed as a limitation, and some of the method described blocks may be combined in a different order to implement the method, or alternative method. Based on the present approaches as described in the context of the example method 400, the input eye image(s) 206 is processed based on the trained stenosis detection model 114 and the CAC detection model 116.

[0074] Further, the above-mentioned method 400 may be implemented in a suitable hardware, computer-readable instructions, or combination thereof. The steps of such method may be performed by either a system under the instruction of machine executable instructions stored on a non- transitory computer readable medium or by dedicated hardware circuits, microcontrollers, or logic circuits. For example, the method may be performed by a detection system, such as detection system 202. In an implementation, the method may be performed under an “as a service” delivery model, where the detection system 202, operated by a provider, receives programmable code. Herein, some examples are also intended to cover non-transitory computer readable medium, for example, digital data storage media, which are computer readable and encode computerexecutable instructions, where said instructions perform some or all the steps of the above-mentioned method.

[0075] In an example, the method 400 may be implemented by the detection system 202 for detecting presence of stenosis and calcification within a coronary artery based on the input eye image(s) 206 of the subject using the trained stenosis detection model 114 and the trained CAC detection model 116, respectively. At block 402, an input eye image pertaining to a fundus of a subject’s eye is obtained. For example, the detection engine 220 within the detection system 202 retrieves the input eye image(s) 206 from the data repository 204. The input eye image(s) 206 may be high-resolution fundus photographs captured using a retinal imaging device, such as a fundus camera or smartphone-based retinal imaging system. These images typically depict the posterior segment of the eye, including vascular structures (arteries and veins), the optic disc, and the macula, which are relevant for assessing systemic vascular health.

[0076] At block 404, values corresponding to a plurality of health markers associated with the subject are obtained. For example, the detection engine 220 retrieves values corresponding to a plurality of health markers, such as health marker data 208, from the data repository 204. In an example, the health marker data 208 includes subject-specific clinical and demographic information that may influence cardiovascular risk. Examples of such health marker data 208 include, but are not limited to, age, gender, height, weight, smoking status, blood pressure values, cholesterol levels, and history of cardiovascular diseases. These markers are used in conjunction with retinal features to improve the accuracy of stenosis detection.

[0077] At block 406, a type of view of the input eye image is determined. For example, the detection engine 220 determines the type of view represented in the input eye image(s) 206. In an example, the type of view indicates the anatomical region captured in the image and may include macula-centered views, optic disc views, or peripheral retinal views. The type of view is determined to ensure that the image contains diagnostically relevant regions necessary for accurate analysis. Based on the determinedtype of view, the detection engine 220 selectively proceeds to either use the input eye image(s) 206 for further processing or inhibit further processing if the image is deemed unsuitable.

[0078] At block 408, a determination is made to check whether the determined type of view is acceptable or not. For example, in cases where the type of view is identified as a macula-centered view or an optic disc view, the detection engine 220 proceeds to use the input eye image(s) 206 for further processing, as these views contain critical vascular and anatomical features, i.e. , proceed through ‘Yes’ path to block 412. On the other hand, if the input eye image(s) 206 corresponds to other view types that do not provide sufficient diagnostic information, the method proceeds through ‘No’ path to block 410.

[0079] At block 410, the input eye image is discarded and a prompt indicating the unsuitability of the same is generated. For example, on determining that the type of view of the input eye image(s) 206 is not suitable, the detection engine 220 inhibits further processing and discards the image. In one example, the detection engine 220 may also generate a prompt on a display device of the detection system 202, indicating that the image is unsuitable for analysis and requesting acquisition of a new image.

[0080] At block 412, a plurality of image enhancement steps are performed on the input eye image. For example, once the input eye image(s) 206 is deemed eligible for further processing, the detection engine 220 performs a plurality of image enhancement steps to improve the visibility and clarity of retinal features. These enhancements may include, but are not limited to, contrast enhancement to highlight blood vessels, sharpness adjustment to improve edge definition, brightness normalization to correct illumination inconsistencies, green channel extraction to emphasize vascular structures, and bilateral filtering to reduce noise while preserving important details. These steps ensure that the image is optimized for accurate feature extraction.

[0081] At block 414, a segmentation process is performed ono the input eye image to extract relevant image segments. For example, the detection engine 220 performs a segmentation process on the input eye image(s) 206 to extract image segments corresponding to vascular structures and anatomical regions of the subject’s eye. In an example, segmentation is performed to isolate arteries, veins, the optic disc, and the optic cup. These segmented regions are used to compute derived metrics such as vessel density, tortuosity, and disc-to-cup ratios, which are critical indicators of systemic vascular health and potential coronary artery stenosis.

[0082] At block 416, at least the input eye image is processed using a stenosis detection model and a CAC detection model to detect presence of stenosis and calcification, respectively, within a coronary artery of the subject. For example, once all relevant data is obtained, derived, and extracted, the detection engine 220 processes the input eye image(s) 206 along with the health marker data 208 and segmented images using the stenosis detection model 114 to detect the presence of stenosis, and using the CAC detection model 116 to detect the presence or extent or location of CAC. In an example, both the models extract values corresponding to a plurality of characteristic attributes, referred to as attribute data 226, from the input eye image(s) 206. The attribute data 226 indicates vascular and morphological details of the fundus of the subject’s eye, including but not limited to disc and cup height and width, vessel density, average vessel width, fractal dimensions, and various tortuosity metrics such as distance tortuosity, squared curvature tortuosity, and tortuosity density. These features are indicative of systemic vascular health and are used to assess the likelihood of narrowing or blockage in the coronary arteries.

[0083] Based on the extracted attribute data 226, the stenosis detection model 114 detects the presence of stenosis, and the CAC detection model 116 detects the presence or extent of CAC in the coronary artery of the subject. In an example, each model may output a binary classification indicating whether the concerned condition is present or not, or a severitylevel such as mild, moderate, or severe. The extent of stenosis or CAC may be quantified based on the correlation between the extracted retinal features and clinically validated indicators, such as percentage narrowing of the artery or calcium scores. This assessment provides a non-invasive means to evaluate coronary artery health and supports early identification of individuals at risk of cardiovascular events.

[0084] In an example, the stenosis detection model 114 and the CAC detection model 116 may output a binary classification (presence or absence), a categorical seventy level (e.g., mild, moderate, severe), or a continuous score, such as stenosis score 228 and the calcium score 230, indicating the extent of narrowing or calcification. These outputs are determined based on the correlation between retinal features and clinically validated indicators for stenosis and CAC.

[0085] At block 418, the stenosis score and the calcium score are obtained which are determined by the stenosis detection model and the CAC detection model. For example, as described above, while the presence of stenosis is a strong indicator of coronary artery disease (CAD), it alone may not confirm the diagnosis. To provide a more comprehensive assessment, the detection engine 220 obtains both the stenosis score 228, generated by the stenosis detection model 114, and the calcium score 230, generated by the CAC detection model 116 or derived from clinical data, corresponding to the subject from the data repository 204 to detect presence of CAD within the subject. In an example, the stenosis score 228 reflects the likelihood or severity of coronary artery stenosis based on retinal features, while the calcium score 230 indicates the extent of calcified plaque within the coronary arteries, a well-established marker of atherosclerotic burden.

[0086] At block 420, based on the calcium score and the stenosis score, presence of coronary artery disease, i.e., CAD is identified within the subject’s cardiovascular system by generating the CAD detection result 232. For example, these scores, along with other relevant health markerdata and attribute data, are provided as input to a dedicated CAD detection model, a separate machine learning model trained to integrate multi-modal information. The CAD detection model is configured to analyze the combined outputs from the stenosis detection model 114 and the CAC detection model 116, as well as additional clinical and demographic features, to determine the overall likelihood or presence of CAD within the subject’s cardiovascular system. The CAD detection model may output a binary classification (presence or absence of CAD), a categorical risk level (e.g., low, moderate, high), or a continuous risk score, based on established clinical thresholds and learned correlations. As a result of such detection based on stenosis and calcium score (228, 230), the detection engine 220 generates the CAD detection result 232 indicating the overall presence, likelihood, or severity of coronary artery disease in the subject. The CAD detection result 232 may be presented as a binary classification, a categorical risk level (e.g., low, moderate, high), or a continuous score, and may be used to support clinical decision-making and early intervention.

[0087] Although the above paragraphs describe the detection of CAD as being based on both stenosis and calcium scores, this should not be considered limiting. The present subject matter may also be implemented using only the stenosis detection model or only the CAC detection model, depending on the diagnostic requirements or model performance. CAD detection using either of these models independently still falls within the scope of the invention and does not deviate from the core principles described herein.

[0088] FIG. 5 illustrates a computing environment 500 implementing a non-transitory computer-readable medium for detecting presence of stenosis within a coronary artery of a subject using a trained stenosis detection model. The computing environment 500 includes processor(s) 502 communicatively coupled to a computer-readable medium 504 through a communication link 506. The processor(s) 502 may have one or moreprocessing resources for fetching and executing computer-readable instructions from the computer-readable medium 504.

[0089] The computer-readable medium 504 may be, for example, an internal memory device or an external memory device. In an example implementation, the communication link 506 may be a network communication link. The processor(s) 502 and the computer-readable medium 504 may also be communicatively coupled to a client device 508 over the network.

[0090] In an example implementation, the computer-readable medium 504 includes a set of computer-readable instructions 510 (referred to as instructions 510) which may be accessed by the processor(s) 502 through the communication link 506. The instructions 510 cause the processor(s) 502 obtain or retrieve an input eye image(s), such as the input eye image(s) 206 from the data repository 204. The input eye image(s) 206 may be high- resolution fundus photographs captured using a retinal imaging device, such as a fundus camera or smartphone-based retinal imaging system. These images typically depict the posterior segment of the eye, including vascular structures (arteries and veins), the optic disc, and the macula, which are relevant for assessing systemic vascular health.

[0091] Once the input eye image is obtained, the instructions 510 may cause the processor(s) 502 to retrieve values corresponding to a plurality of health markers, such as health marker data 208, from the data repository 204. In an example, the health marker data 208 includes subject-specific clinical and demographic information that may influence cardiovascular risk. Examples of such health marker data 208 include, but are not limited to, age, gender, height, weight, smoking status, blood pressure values, cholesterol levels, and history of cardiovascular diseases. These markers are used in conjunction with retinal features to improve the accuracy of stenosis detection.

[0092] Thereafter, the instructions 510 may cause the processor(s) 502 to determine the type of view represented in the input eye image(s) 206. Inan example, the type of view indicates the anatomical region captured in the image and may include macula-centered views, optic disc views, or peripheral retinal views. The type of view is determined to ensure that the image contains diagnostically relevant regions necessary for accurate analysis. Based on the determined type of view, the detection engine 220 selectively proceeds to either use the input eye image(s) 206 for further processing or inhibit further processing if the image is deemed unsuitable.

[0093] In cases where the type of view is identified as a maculacentered view or an optic disc view, the instructions 510 causes the processor(s) 502 to proceed to use the input eye image(s) 206 for further processing, as these views contain critical vascular and anatomical features. On the other hand, if the image corresponds to other view types that do not provide sufficient diagnostic information, the instructions 510 causes the processor(s) 502 to inhibit further processing and discards the image. In one example, a prompt on a display device of the detection system 202 may also be generated, indicating that the image is unsuitable for analysis and requesting acquisition of a new image.

[0094] Returning to the present example, once the input eye image(s) 206 is deemed eligible for further processing, the instructions 510 causes the processor(s) 502 to perform a plurality of image enhancement steps to improve the visibility and clarity of retinal features. These enhancements may include, but are not limited to, contrast enhancement to highlight blood vessels, sharpness adjustment to improve edge definition, brightness normalization to correct illumination inconsistencies, green channel extraction to emphasize vascular structures, and bilateral filtering to reduce noise while preserving important details. These steps ensure that the image is optimized for accurate feature extraction.

[0095] Thereafter, the instructions 510 causes the processor(s) 502 to perform a segmentation process on the input eye image(s) 206 to extract image segments corresponding to vascular structures and anatomical regions of the subject’s eye. In an example, segmentation is performed toisolate arteries, veins, the optic disc, and the optic cup. These segmented regions are used to compute derived metrics such as vessel density, tortuosity, and disc-to-cup ratios, which are critical indicators of systemic vascular health and potential coronary artery stenosis.

[0096] Once all relevant data is obtained, derived, and extracted, the instructions 510 causes the processor(s) 502 to process the input eye image(s) 206 along with the health marker data 208 and segmented images using the stenosis detection model 114 to detect the presence of stenosis, and using the CAC detection model 116 to detect the presence or extent or location of CAC. In an example, both the models extracts values corresponding to a plurality of characteristic attributes, referred to as attribute data 226, from the input eye image(s) 206. The attribute data 226 indicates vascular and morphological details of the fundus of the subject’s eye, including but not limited to disc and cup height and width, vessel density, average vessel width, fractal dimensions, and various tortuosity metrics such as distance tortuosity, squared curvature tortuosity, and tortuosity density. These features are indicative of systemic vascular health and are used to assess the likelihood of narrowing or blockage in the coronary arteries.

[0097] Based on the extracted attribute data 226, the stenosis detection model 114 detects the presence of stenosis, and the CAC detection model 116 detects the presence or extent of CAC in the coronary artery of the subject. In an example, each model may output a binary classification indicating whether the concerned condition is present or not, or a severity level such as mild, moderate, or severe. The extent of stenosis or CAC may be quantified based on the correlation between the extracted retinal features and clinically validated indicators, such as percentage narrowing of the artery or calcium scores. This assessment provides a non-invasive means to evaluate coronary artery health and supports early identification of individuals at risk of cardiovascular events.

[0098] In an example, the stenosis detection model 114 and the CAC detection model 116 may output a binary classification (presence or absence), a categorical seventy level (e.g., mild, moderate, severe), or a continuous score, such as stenosis score 228 and the calcium score 230, indicating the extent of narrowing or calcification. These outputs are determined based on the correlation between retinal features and clinically validated indicators for stenosis and CAC.

[0099] As described above, while the presence of stenosis is a strong indicator of coronary artery disease (CAD), it alone may not confirm the diagnosis. In such cases, the instructions 510 causes the processor(s) 502 to obtain both the stenosis score 228, generated by the stenosis detection model 114, and the calcium score 230, generated by the CAC detection model 116 or derived from clinical data, corresponding to the subject from the data repository 204 to detect presence of CAD within the subject. In an example, the stenosis score 228 reflects the likelihood or seventy of coronary artery stenosis based on retinal features, while the calcium score 230 indicates the extent of calcified plaque within the coronary arteries, a well-established marker of atherosclerotic burden.

[0100] In one example, these scores, along with other relevant health marker data and attribute data, are provided as input to a dedicated CAD detection model, a separate machine learning model trained to integrate multi-modal information. The CAD detection model is configured to analyze the combined outputs from the stenosis detection model 114 and the CAC detection model 116, as well as additional clinical and demographic features, to determine the overall likelihood or presence of CAD within the subject’s cardiovascular system. The CAD detection model may output a binary classification (presence or absence of CAD), a categorical risk level (e.g., low, moderate, high), or a continuous risk score, based on established clinical thresholds and learned correlations.

[0101] Although examples for the present disclosure have been described in language specific to structural features and / or methods, it is tobe understood that the appended claims are not necessarily limited to the specific features or methods described. Rather, the specific features and methods are disclosed and explained as examples of the present disclosure.

Claims

I / We Claim:1 . A system comprising: a processor; and a coronary artery disease (CAD) detection engine coupled to the processor, wherein the CAD detection engine is to: obtain an input eye image pertaining to a fundus of a subject’s eye; and process at least the input eye image using a stenosis detection model, wherein the stenosis detection model is trained based on training data comprising at least a training fundus image and a reference stenosis indicator indicating status of stenosis corresponding to the training fundus image, and wherein the stenosis is indicative of one of narrowing and blockage of a coronary artery, the stenosis detection model is to: extract values corresponding to a plurality of characteristic attributes from the input eye image indicating vascular and morphological details of the fundus of the subject’s eye; and based on the extracted values, detect presence of stenosis in the coronary artery of the subject.

2. The system as claimed in claim 1 , wherein CAD detection engine is to: obtain values corresponding to a plurality of health markers associated with the subject, wherein the plurality of health markers comprises age, gender, height, weight, smoking status, blood pressure value, cholesterol levels, and history of cardiovascular diseases; and process the values corresponding to the plurality of health markers along with the input eye image using the stenosis detection model to detect presence of stenosis.

3. The system as claimed in claim 1 , wherein the CAD detection engine is to:process at least the input eye image using a CAC detection model, wherein the CAC detection model is trained based on training data comprising at least a training fundus image and a reference CAC indicator indicating status of calcification corresponding to the training fundus image, and wherein the CAC is indicative of presence of calcium plaque within a coronary artery, the CAC detection model is to: extract values corresponding to a plurality of characteristic attributes from the input eye image indicating vascular and morphological details of the fundus of the subject’s eye; and based on the extracted values, detect presence of calcification in the coronary artery of the subject.

4. The system as claimed in claim 1 , wherein the CAD detection engine is to: obtain a stenosis score and a calcium score as output from the stenosis detection model and the CAC detection model; and based on the stenosis score and the calcium score, identify presence of coronary artery disease (CAD) within the subject’s cardiovascular system.

5. The system as claimed in claim 1 , wherein the plurality of characteristic attributes comprises one or more of a disc & cup height, fractal dimensions, a vessel density, an average width of vessels, distance tortuosity, squared curvature tortuosity, and tortuosity density.

6. The system as claimed in claim 1 , wherein the CAD detection engine is to: determine a type of view of the input eye image, wherein the type of view indicates an anatomical region captured in the input eye image;based on the determined type of view, selectively one of enable further processing and inhibit further processing of the input eye image; and in response to inhibiting, generate a prompt indicating unsuitability of the image for analysis and requests acquisition of a new image.

7. The system as claimed in claim 1 , wherein the CAD detection engine is to: perform a plurality of image enhancement steps on the input eye image comprising contrast enhancement, sharpness adjustment, brightness adjustment, green channel extraction, and bilateral filtering; and perform a segmentation process on the input eye image to extract image segments corresponding to vascular structures, comprising arteries and veins, and anatomical regions, comprising the optic disc and the optic cup.

8. The system as claimed in claim 1 , wherein the reference stenosis indicator comprises a combination of clinical and imaging-based parameters comprising calcium scores, invasive stenosis readings and non-inva- sive stenosis readings.

9. A method comprising: obtaining training data comprising at least training fundus image and a reference stenosis indicator indicating status of stenosis corresponding to the training fundus image; and training a stenosis detection model based on the training data, wherein the stenosis detection model, when trained, is to: extract values corresponding to a plurality of characteristic attributes indicating vascular and morphological details from an input eye image; and based on the extracted values, detect presence of stenosis in a coronary artery of the subject.

10. The method as claimed in claim 9, wherein the method comprises: training a CAC detection model based on the training data, wherein the CAC detection model, when trained, is to: extract values corresponding to a plurality of characteristic attributes indicating vascular and morphological details from an input eye image; and based on the extracted values, detect presence of calcification in a coronary artery of the subject.11 . The method as claimed in claim 9, wherein the training data further comprises training values corresponding to the plurality of characteristic attributes corresponding to the training fundus image, wherein the plurality of characteristic attributes comprises disc & cup height, fractal dimensions, vessel density, average width of vessels, distance tortuosity, squared curvature tortuosity, and tortuosity density.

12. The method as claimed in claim 9, wherein the method comprises: performing a plurality of image enhancement steps on the training fundus image, wherein the plurality of image enhancement steps comprising contrast enhancement, sharpness adjustment, brightness adjustment, green channel extraction, and bilateral filtering; and performing a segmentation process on the training fundus image to extract image segments corresponding to vascular structures, comprising arteries and veins, and anatomical regions, comprising the optic disc and the optic cup.

13. The method as claimed in claim 9, wherein the training data further comprises training values corresponding to a plurality of health markers corresponding to the training fundus image, wherein the plurality of healthmarkers comprises age, gender, height, weight, smoking status, blood pressure value, cholesterol levels, history of cardiovascular diseases.

14. The method as claimed in claim 9, wherein the reference stenosis indicator and the reference CAC indicator comprises a combination of clinical and imaging-based parameters comprising the invasive readings and non-invasive readings corresponding to the stenosis and CAC.

15. A non-transitory computer-readable medium comprising instructions, the instructions being executable by a processing resource to: obtain an input eye image pertaining to a fundus of a subject’s eye; and process the input eye image using a stenosis detection model, wherein the stenosis detection model is trained based on training data comprising at least a training fundus image and a reference stenosis indicator indicating status of stenosis corresponding to the training fundus image, and wherein the stenosis is indicative of one of narrowing and blockage of a coronary artery, the stenosis detection model is to: extract values corresponding to a plurality of characteristic attributes from the input eye image indicating vascular and morphological details of the fundus of the subject’s eye; and based on the extracted values, detect presence of stenosis in the coronary artery of the subject.

16. The non-transitory computer-readable medium as claimed in claim 15, wherein the instructions being executable by the processing resource to: obtain values corresponding to a plurality of health markers associated with the subject, wherein the plurality of health markers comprises age, gender, height, weight, smoking status, blood pressure value, cholesterol levels, and history of cardiovascular diseases; andprocess the values corresponding to the plurality of health markers along with the input eye image using the stenosis detection model to detect presence of stenosis.

17. The non-transitory computer-readable medium as claimed in claim 15, wherein the instructions being executable by the processing resource to: process at least the input eye image using a CAC detection model, wherein the CAC detection model is trained based on training data comprising at least a training fundus image and a reference CAC indicator indicating status of calcification corresponding to the training fundus image, and wherein the CAC is indicative of presence of calcium plaque within a coronary artery, the CAC detection model is to: extract values corresponding to a plurality of characteristic attributes from the input eye image indicating vascular and morphological details of the fundus of the subject’s eye; and based on the extracted values, detect presence of calcification in the coronary artery of the subject.

18. The non-transitory computer-readable medium as claimed in claim 15, wherein the plurality of characteristic attributes comprises one or more of a disc & cup height, fractal dimensions, a vessel density, an average width of vessels, distance tortuosity, squared curvature tortuosity, and tortuosity density.

19. The non-transitory computer-readable medium as claimed in claim 15, wherein the instructions being executable by the processing resource to:perform a plurality of image enhancement steps on the input eye image comprising contrast enhancement, sharpness adjustment, brightness adjustment, green channel extraction, and bilateral filtering; and perform a segmentation process on the input eye image to extract image segments corresponding to vascular structures, comprising arteries and veins, and anatomical regions, comprising the optic disc and the optic cup.

20. The non-transitory computer-readable medium as claimed in claim 15, wherein the instructions being executable by the processing resource to: obtain a stenosis score and a calcium score as output from the stenosis detection model and the CAC detection model; and based on the stenosis score and the calcium score, identify presence of coronary artery disease (CAD) within the subject’s cardiovascular system.

Images