System and method for detecting opthalmic conditions using machine learning

Abstract

A system for detecting retinal detachment is provided herein. The system includes an imaging device configured to capture a retinal fundus image of an individual, a processor, and a memory storing instructions. When executed by the processor, the instructions cause the system to preprocess the retinal fundus image to remove extraneous background information and enhance retinal features, input the preprocessed retinal fundus image into a trained machine learning model, generate a determination of whether the retinal fundus image indicates presence of retinal detachment using the trained machine learning model, and output the determination to a user interface. The machine learning model is trained on a set of labeled retinal fundus images to classify images as indicating presence or absence of retinal detachment.

Description

SYSTEM AND METHOD FOR DETECTING OPTHALMIC CONDITIONS USINGMACHINE LEARNINGCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority from U.S. Provisional Application Serial No. 63 / 649,957 filed on May 21, 2024, which is incorporated herein by reference in its entirety.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] Not Applicable.INTRODUCTION

[0003] Retinal detachment is a serious eye condition that can lead to permanent vision loss if not promptly diagnosed and treated. Traditional methods for detecting retinal detachment rely heavily on manual examination by trained ophthalmologists using specialized equipment. However, these approaches have limitations in terms of accessibility7, scalability, and consistency.

[0004] Existing techniques for retinal examination include indirect ophthalmoscopy, where an ophthalmologist examines the retina through a handheld lens, and retinal fundus imaging, which captures photographs of the back of the eye. While these methods can be effective when performed by skilled practitioners, they face challenges in widespread implementation, especially in areas with limited access to eye care specialists.

[0005] Automated systems for analyzing retinal images have emerged as a potential solution to improve the efficiency and reach of retinal health screening. Some approaches utilize basic image processing algorithms to detect abnormalities in fundus photographs. However, these systems often struggle with the complexity and variability of retinal features, leading to suboptimal accuracy and reliability.

[0006] More recently, machine learning techniques have been applied to retinal image analysis. Early machine learning models showed promise in detecting certain retinal conditions, but they frequently required extensive feature engineering and had difficulty generalizing across diverse patient populations and imaging conditions.

[0007] Deep learning, a subset of machine learning utilizing neural networks with multiple layers, has demonstrated impressive results in various medical imaging tasks. However, its application to retinal detachment detection presents unique challenges. Retinal detachment can manifest in subtle ways that are difficult to discern from normal variations or other retinal conditions. Additionally, the limited availability7of large, well -annotated datasets of retinal detachment images has hindered the development of robust deep learning models for this specific task.

[0008] Furthermore, existing automated retinal analysis systems often operate as “black boxes,” providing predictions without clear explanations. This lack of interpretability7can be problematic in clinical settings where understanding the reasoning behind a diagnosis is crucial for treatment planning and patient communication.

[0009] Another limitation of current approaches is the reliance on high-quality, standardized images captured under ideal conditions. In real-world scenarios, image quality can vary7significantly due to factors such as patient cooperation, camera positioning, and lighting conditions. Systems that perform well on curated datasets may struggle when faced with the variability encountered in practical applications.

[0010] Lastly , many existing solutions focus solely on image analysis without considering the broader context of patient care. Factors such as medical history7, symptoms, and risk factors play a crucial role in accurate diagnosis and management of retinal detachment. Integrating these diverse data sources with image analysis remains a challenge for automated systems.

[0011] Given these limitations, there is a clear need for improved methods and systems that can accurately detect retinal detachment from fundus images while addressing the challenges of real-world variability7, interpretability, and holistic patient assessment.

[0012] Several challenges persist in the prior art of retinal detachment detection. Traditional methods rely7heavily on manual examination by skilled ophthalmologists, limiting accessibility and scalability, especially in areas with limited access to eye care specialists. Existing automated systems often struggle with the complexity and variability of retinal features, leading to suboptimal accuracy and reliability. While machine learning approaches have shown promise, they frequently require extensive feature engineering and have difficulty generalizing across diverse patient populations and imaging conditions. Furthermore, many current systems operate as “black boxes,” lacking interpretability crucial for clinical decision-making. The reliance on high-quality, standardized images poses challenges in real-world scenarios where image quality can vary significantly. Lastly, theintegration of diverse data sources such as medical history, symptoms, and risk factors with image analysis remains a significant challenge. These unresolved issues underscore the pressing demand for improved methods and systems that can accurately detect retinal detachment while addressing real-world variability, interpretability, and holistic patient assessment.SUMMARY

[0013] The present disclosure provides systems, methods, and non-transitory computer-readable media for detecting ophthalmic conditions using machine learning techniques. In some aspects, a system for detecting ophthalmic conditions such as retinal detachment, Central Retinal Artery Occlusion (CRAO). Endophthalmitis, and Papilledema are provided. The system may include an imaging device configured to capture a retinal fundus image of an individual, a processor, and a memory storing instructions. When executed, these instructions may cause the system to preprocess the retinal fundus image to enhance retinal features, input the preprocessed image into a trained machine learning model, generate a determination of whether the image indicates presence of retinal detachment, and output the determination to a user interface.

[0014] In some implementations, the imaging device may comprise a portable device with an adjustable magnification lens configured to align with a camera of an electronic device. The portable device may include a lens slider to adjust the position of the magnification lens and a device holder to secure the electronic device in alignment with the magnification lens.

[0015] The trained machine learning model may be configured to classify the retinal fundus image as indicating the presence of one or more specific types of retinal detachment, such as rhegmatogenous retinal detachment, exudative retinal detachment, and tractional retinal detachment. In some cases, the system may be further configured to receive an ophthalmologist analysis of the retinal fundus image and update the trained machine learning model based on this analysis.

[0016] The system may also be capable of flagging cases that require immediate intervention for urgent ophthalmologist review, such as macula-on retinal detachments. The preprocessing of the retinal fundus image may involve removing extraneous background information. In some implementations, the determination generated by the system may include a probability score indicating the likelihood of retinal detachment.

[0017] Various types of machine learning models may be employed, includingconvolutional neural networks, ResNet, MobileNet, EfficientNet, DenseNet, and Inception- ResNet. The trained machine learning model may be configured to generate a heatmap highlighting regions of the retinal fundus image indicative of retinal detachment.Additionally, the model may classify the retinal fundus image into multiple categories, including normal, early-stage retinal detachment, and advanced retinal detachment.

[0018] The disclosure also presents methods for detecting retinal detachment, which may involve capturing a retinal fundus image, preprocessing the image, inputting it into a trained machine learning model, generating a determination, and outputting the determination to a user interface. These methods may also include receiving ophthalmologist analysis and updating the trained machine learning model based on this analysis.

[0019] In some aspects, the disclosure provides a computer-implemented method for detecting ophthalmic diseases, disorders, or conditions. This method may involve generating a machine learning classifier, obtaining an ophthalmic image, evaluating the image using the classifier, and providing a determination to the individual or a third part}'. The determination may have a sensitivity and specificity of at least 90% when tested against an independent data set of at least 100 sample images.

[0020] The disclosure further describes computer-implemented systems that may include a server for generating a machine learning classifier, an electronic device with a camera, a portable device with an imaging component, and a computer program for capturing and analyzing ophthalmic images. Another system configuration may include a medical imaging device coupled to an electronic device for capturing and analyzing ophthalmic images.

[0021] These systems and methods may be applied to various ophthalmic diseases, with a particular focus on retinal conditions such as retinal detachment. The systems may provide recommendations on areas of concern in the ophthalmic image and suggestions for positioning the individual to potentially slow disease progression. The determination of the retinal condition may be provided through various interfaces, including web interfaces, mobile interfaces, electronic devices, or communicated via medical personnel.

[0022] These and other features, aspects and advantages of the present teachings will become better understood with reference to the following description, examples and appended claims.DRAWINGS

[0023] Those of skill in the art will understand that the drawings, described below,are for illustrative purposes only. The drawings are not intended to limit the scope of the present teachings in any way.

[0024] FIG. 1 illustrates a process for preparing a retinal fundus image, optimizing it for subsequent analysis and classification.

[0025] FIG. 2 depicts a method for training a machine learning model, enabling accurate retinal detachment classification from fundus images.

[0026] FIG. 3 demonstrates the application of a trained model for analyzing retinal fundus images, facilitating efficient diagnosis.

[0027] FIG. 4 shows a process for updating the machine learning model, incorporating expert ophthalmologist analysis to enhance accuracy.

[0028] FIG. 5 presents a comprehensive method for detecting retinal detachment, integrating model training, deployment, and continuous improvement.

[0029] FIG. 6 illustrates an apparatus for capturing and analyzing retinal fundus images, designed to facilitate accurate image acquisition and processing.

[0030] FIG. 7 illustrates a computer system for use in the present invention.DETAILED DESCRIPTION

[0031] All patents, applications, published applications and other publications cited herein are incorporated by reference in their entirety. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the invention belongs.

[0032] Any methods, devices and materials similar or equivalent to those described herein can be used in the practice of this invention. The following definitions are provided to facilitate understanding of certain terms used frequently herein and are not meant to limit the scope of the present disclosure. In the event that there is a plurality of definitions for a term herein, those in this section prevail unless stated otherwise. Headings used herein are for organizational purposes only and in no way limit the invention described herein.

[0033] Abbreviations and Definitions

[0034] To facilitate understanding of the invention, a number of terms and abbreviations as used herein are defined below as follows:

[0035] Opthalmic condition: As used herein, the term “ophthalmic condition"’ refers to retinal detachment, Central Retinal Artery Occlusion (CRAO). Endophthalmitis, and Papilledema. The term also includes ophthalmic conditions predicted by those of skill in theart to be detected by the systems of the present disclosure.

[0036] Retinal detachment: As used herein, the term "retinal detachment’7refers to a separation of the retina from the underlying layers of the eye, which can lead to vision loss if left untreated. The term “normal” in relation to retinal detachment refers to a retina that is fully attached to the underlying layers of the eye, with no signs of separation or detachment visible in imaging studies or clinical examination. The term “early-stage retinal detachment" may refer to a condition where there is a small, localized separation of the retina from the underlying layers, typically affecting less than 25% of the retinal area. This stage may be characterized by the presence of retinal tears or holes, fluid accumulation beneath the retina, or the initial lifting of a small portion of the retina. The term “advanced retinal detachment” may refer to a more extensive separation of the retina from the underlying layers, typically affecting more than 25% of the retinal area. This stage may be characterized by a larger area of retinal detachment, potentially involving the macula, and may be associated with more severe visual symptoms. Advanced retinal detachment may also include cases where the detachment has been present for a longer duration or where there are complications such as proliferative vitreoretinopathy.

[0037] Ophthalmic image: As used herein, the term “ophthalmic image” refers to any visual representation of the eye or its structures, captured through various imaging modalities for diagnostic, therapeutic, or research purposes. This includes, but is not limited to, retinal fundus images and ultra-widefield fundus images, ultra-widefield fundus images, optical coherence tomography (OCT) scans, fluorescein angiography, indocyanine green angiography, scanning laser ophthalmoscopy images, anterior segment photographs, comeal topography maps, b-scan ultrasound, and any other digital or analog visual representation of ocular tissues. Ophthalmic images may be captured using specialized ophthalmic imaging equipment in clinical settings or through adapted consumer devices with appropriate attachments. These images may be in various formats including two-dimensional photographs, three-dimensional reconstructions, or time-series captures, and may represent visible light reflections, fluorescence emissions, or other electromagnetic spectrum visualizations of ocular structures.

[0038] Retinal fundus image: As used herein, the term “retinal fundus image” refers to a photograph of the back of the eye, including the retina, optic disc, macula, and blood vessels.

[0039] Ultra-widefield fundus (UWF) image: As used herein, the term “ultra- widefield fundus (UWF) image” refers to a high-resolution photograph capturing a widerview of the retina compared to standard fundus images, typically covering up to 200 degrees of the retinal surface.

[0040] Rhegmatogenous retinal detachment: As used herein, the term “rhegmatogenous retinal detachment” refers to a type of retinal detachment caused by a tear or break in the retina, allowing fluid to accumulate underneath and separate the retina from the underlying tissue.

[0041] Non-rhegmatogenous retinal detachment: As used herein, the term “non- rhegmatogenous retinal detachment” is broadly defined to include exudative retinal detachment, caused by fluid accumulation without a retinal break, and fractional retinal detachment, caused by fibrous or fibrovascular tissue pulling the retina away from its normal position. Those of skill in the art will recognize other tears that are not complete, and are detectable using the system and apparatus of the present invention.

[0042] System and Method for Detecting Retinal Detachment Using Machine Learning

[0043] The systems and methods described herein relate generally to obtaining retinal fundus images and using machine learning for detecting retinal detachment in the retinal fundus images. This approach represents a novel and non-obvious solution to the challenge of early detection of retinal detachment, a serious ocular condition that can lead to vision loss if left untreated.

[0044] Traditionally, detection of retinal detachment has relied on manual examination by trained ophthalmologists, which can be time-consuming and subject to human error. The use of machine learning algorithms to analyze retinal fundus images may provide a more efficient and potentially more accurate method for detecting retinal detachment.

[0045] The invention described herein may address several technical problems in the field of ophthalmology. First, it may provide a means to rapidly screen large numbers of retinal images, potentially increasing the speed and scale of retinal detachment detection. Second, the use of machine learning algorithms may allow for the detection of subtle signs of retinal detachment that might be missed by human observers. Third, this approach may help address the shortage of trained ophthalmologists in many parts of the world by providing a tool that can assist in the initial screening process.

[0046] The systems and methods described may have broad applicability in the field of ophthalmology. While the primary focus is on the detection of retinal detachment, theunderlying principles and techniques may be adaptable to the detection of other retinal conditions. This versatility may enhance the overall uti 11 ty and impact of the invention in clinical practice.

[0047] In some cases, the machine learning algorithms used in this invention may be trained on large datasets of retinal fundus images, allowing them to leam patterns and features associated with retinal detachment. This training process may enable the algorithms to generalize their learning to new, unseen images, potentially providing robust and reliable detection capabilities.

[0048] The invention may also incorporate various types of retinal imaging techniques, including standard fundus photography and ultra-widefield imaging. This flexibility in image input may allow the system to be used with a variety of existing ophthalmic imaging equipment, potentially facilitating its adoption in diverse clinical settings.

[0049] Those skilled in the art will recognize that various modifications and variations can be made without departing from the spirit and scope of the invention. The specific implementations and applications described in the following sections should be considered as non-limiting examples of how the invention may be realized in practice.

[0050] Image Preparation Process

[0051] The image preparation process may involve several steps to obtain a clean retinal fundus image suitable for analysis. Figure 1 illustrates an exemplary process for preparing a retinal fundus image.

[0052] In some cases, the process may begin with a step 110 of obtaining a retinal image using an image capturing device. The retinal image obtained in step 110 may be a standard fundus image or an ultra widefield fundus (UWF) image. UWF images may provide a wider view of the retina, potentially capturing more peripheral areas that could be relevant for detecting retinal detachment.

[0053] In some cases, optical coherence tomography (OCT) may be utilized as a complementary imaging technique to confirm or further evaluate suspected retinal detachment. OCT is a non-invasive imaging method that provides high-resolution, cross- sectional images of the retina and underlying structures. This technology may offer several advantages in the context of retinal detachment detection and assessment.

[0054] OCT may use low-coherence light to capture detailed images of the retinal layers, allowing for visualization of subtle structural changes that may not be apparent infundus photographs. The cross-sectional nature of OCT images may enable precise localization of retinal detachments, including the ability to distinguish between different types of detachments such as rhegmatogenous, tractional, or exudative.

[0055] In some implementations, OCT may be integrated into the retinal detachment detection workflow7. For instance, cases flagged as suspicious by the machine learning analysis of fundus images may be recommended for follow-up OCT imaging. This multi-modal approach may enhance the overall accuracy of the diagnosis and provide additional information for treatment planning.

[0056] OCT may also be valuable in assessing the extent of retinal detachment, including whether the macula is involved (macula-on vs. macula-off detachment). This distinction may be critical for determining the urgency of treatment and predicting visual outcomes. Additionally, OCT may help in identifying associated pathologies such as vitreomacular traction or epiretinal membranes that could complicate the management of retinal detachment.

[0057] The integration of OCT data into the machine learning model may further improve its performance. In some cases, the system may be trained on both fundus images and corresponding OCT scans, allowing it to correlate features across different imaging modalities. This multi-modal learning approach may enhance the model's ability to detect and characterize retinal detachments accurately.

[0058] Furthermore, OCT may play a role in post-treatment follow-up, allowing for precise monitoring of retinal reattachment and recovery. The high-resolution imaging provided by OCT may enable detection of subtle complications or incomplete reattachment that might not be visible through other examination methods.

[0059] While OCT may provide valuable additional information, it may not always be necessary or available in all clinical settings. The machine learning system described herein may be designed to function effectively with fundus images alone, with OCT serving as an optional, complementary tool for cases where further confirmation or detailed structural analysis is required.

[0060] Following image capture, the system may employ image quality assessment techniques to ensure that the obtained retinal fundus image meets certain quality standards before proceeding with further analysis. This quality assessment step may involve evaluating various aspects of the image, such as focus, contrast, illumination, and the presence of artifacts.

[0061] In some implementations, the system may utilize a machine learning-basedimage quality assessment model. This model may be trained on a diverse set of retinal fundus images, labeled with quality scores by expert graders. The model may output a quality score for each input image, ranging from 0 to 1, where higher scores indicate better image quality.

[0062] The system may define a quality threshold, for example, 0.7, above which an image is considered suitable for analysis. Images that meet or exceed this threshold may proceed to the next steps of the image preparation process. For images that fall below the threshold, the system may provide feedback to the operator, suggesting adjustments to improve image quality. This feedback may include recommendations for refocusing, adjusting illumination, or repositioning the imaging device.

[0063] In some cases, the quality assessment may be more granular, evaluating specific regions of interest within the retinal fundus image. For instance, the system may assign higher importance to the quality of the macular region and the peripheral retina, as these areas may be particularly relevant for detecting retinal detachment.

[0064] If an image fails to meet the quality' threshold after multiple capture attempts, the system may flag the case for manual review by a trained technician or ophthalmologist. This ensures that potentially important cases are not missed due to technical limitations in image acquisition.

[0065] For images that pass the quality' threshold, the system may proceed with the subsequent steps of the image preparation process. This may include step 120, where extraneous background information is removed from the image, as illustrated in Figure 1.

[0066] Following image capture, a step 120 may involve removing extraneous background information from the image. This step may help isolate the relevant retinal features by eliminating non-retinal elements that could interfere with subsequent analysis.

[0067] The process may continue with a step 130 of further cleaning the image to focus on the retina. This step may involve enhancing contrast, adjusting brightness, or applying filters to improve the visibility of retinal structures.

[0068] In some implementations, the system may incorporate a quality' control mechanism to ensure that only retinal fundus images meeting certain quality standards are processed for retinal detachment detection. This quality control step may involve evaluating various image characteristics and discarding or flagging images that do not meet predefined criteria.

[0069] The system may employ a machine learning-based quality assessment model trained on a diverse set of retinal fundus images with expert-assigned quality scores. This model may evaluate factors such as focus, contrast, illumination, color balance, and thepresence of artifacts or obstructions. The model may output a quality score for each input image, typically ranging from 0 to 1, where higher scores indicate better image quality.

[0070] A quality threshold may be established, for example, 0.7, below which images are considered unsuitable for analysis. Images falling below this threshold may be automatically discarded from the analysis pipeline. In some cases, the system may store these discarded images in a separate database for potential future review or for use in improving the image capture process.

[0071] For images that fall slightly below the quality threshold, the system may attempt automated enhancement techniques such as contrast adjustment, noise reduction, or sharpening. If these enhancements successfully bring the image quality above the threshold, the enhanced image may proceed to the next stage of analysis. Otherwise, it may be discarded.

[0072] In some implementations, the system may employ a tiered quality assessment approach. Images scoring above a high threshold (e.g., 0.9) may proceed directly to analysis. Images scoring between a lower and upper threshold (e.g., 0.7 to 0.9) may be flagged for manual review by a trained technician before proceeding. Images sconng below the lower threshold may be automatically discarded.

[0073] The system may also track the frequency of discarded images and provide feedback to improve the image acquisition process. For example, if a particular imaging device or operator consistently produces low-quality images, the system may generate alerts or recommendations for equipment maintenance or additional operator training.

[0074] In cases where critical diagnostic decisions are being made, the system may require multiple high-quality images of the same retina before proceeding with analysis. This redundancy may help ensure the reliability’ of the retinal detachment detection process.

[0075] By implementing these quality control measures and discarding substandard images, the system may help maintain the accuracy and reliability of the retinal detachment detection process, potentially reducing false positives and negatives that could arise from analyzing poor-quality images.

[0076] In some cases, a step 140 may be performed to remove any glare or other obstructions from the image. Glare can often occur in retinal imaging due to reflections from the eye’s surfaces or optical components of the imaging system. Removing these artifacts may improve the quality of the image for subsequent analysis.

[0077] The specific techniques used in each step may vary depending on the characteristics of the input image and the requirements of the subsequent analysis. Forexample, different image processing algorithms may be applied to standard fundus images versus UWF images to account for their distinct characteristics.

[0078] In some cases, the image preparation process may be automated using computer vision techniques. In other cases, manual intervention by a trained technician may be required for optimal results. The choice between automated and manual processing may depend on factors such as the volume of images to be processed, the complexity of the images, and the desired level of accuracy.

[0079] The image preparation process may incorporate both automated and manual techniques to achieve optimal results. In some implementations, the system may employ a hybrid approach that combines the efficiency of automated processing with the precision of human expertise.

[0080] Automated processing may begin immediately after image acquisition. Computer vision algorithms may be utilized to perform initial quality assessments, identifying issues such as poor focus, inadequate illumination, or excessive artifacts. These algorithms may analyze various image parameters, including contrast, sharpness, and color balance, to generate a preliminary quality score.

[0081] For images that meet a predefined quality threshold, the automated system may proceed with background removal and image enhancement. This may involve segmentation techniques to isolate the retinal area of interest, followed by contrast adjustment and noise reduction algorithms. Machine learning models trained on large datasets of retinal images may be employed to identify and remove common artifacts such as lens reflections or eyelash shadows.

[0082] In cases where the automated system encounters difficulties or produces results below a certain confidence threshold, the process may seamlessly transition to manual intervention. A trained technician or ophthalmologist may be alerted to review the image and make necessary adjustments. This human-in-the-loop approach may be particularly valuable for complex cases or when dealing with unusual retinal presentations.

[0083] The manual component of the process may involve specialized image editing software designed for medical imaging. Technicians may use tools such as brush-based selection for precise artifact removal, localized contrast enhancement to highlight subtle retinal features, and manual annotation to mark areas of interest or concern. In some cases, the manual process may also include capturing additional images or adjusting imaging parameters to obtain a higher quality result.

[0084] To facilitate efficient manual processing, the system may provide anintuitive user interface that allows technicians to quickly navigate between automated suggestions and manual controls. This interface may display side-by-side comparisons of the original and processed images, enabling easy verification of the automated results and streamlined manual corrections when necessary.

[0085] The system may also incorporate a feedback loop between the manual and automated processes. When manual corrections are made, the system may log these adjustments and use them to refine the automated algorithms over time. This continuous learning approach may help improve the accuracy and efficiency of the automated components, potentially reducing the need for manual intervention in future cases.

[0086] In some implementations, the system may employ a tiered processing approach based on the complexity and criticality of the case. Routine screenings may rely more heavily on automated processing to handle high volumes efficiently, while suspected cases of retinal detachment or other serious conditions may automatically trigger a more thorough manual review process.

[0087] The combination of automated and manual processing may also extend to the quality control stage. While automated metrics may flag potentially problematic images, final approval for analysis may require human verification in certain cases. This dual-layer quality7control may help ensure that only images of sufficient quality7proceed to the retinal detachment detection phase.

[0088] By integrating both automated and manual processes, the system may achieve a balance between efficiency and accuracy. The automated components may handle the bulk of routine processing tasks, while the option for manual intervention provides flexibility7to address complex or ambiguous cases. This hybrid approach may allow the system to adapt to various clinical settings and image quality scenarios, potentially improving the overall reliability of retinal detachment detection.

[0089] In some implementations, the machine learning model may be trained to identify specific features and objects in retinal images that are indicative of retinal detachment. These features may be present in both fundus images and optical coherence tomography (OCT) scans.

[0090] For fundus images, the model may be configured to detect Shafer's sign, which appears as a demarcation line between the attached and detached retina. The model may also identify areas of retinal corrugation, which can manifest as wavy or folded regions in the image.

[0091] The system may be trained to recognize retinal tears or holes, which canappear as dark spots or irregularities in the fundus image. Additionally, the model may detect lifted or folded areas of the retina, often accompanied by fluid accumulation underneath, which can present as elevated or shadowed regions in the image.

[0092] In OCT scans, the machine learning model may be designed to identify specific features associated with retinal detachment. These may include the presence of a hyperreflective layer corresponding to the detached retina, as well as areas where the retina appears to dip or tent into the subretinal space. The model may also detect subretinal septa, which can appear as thin lines extending between the detached retina and the underlying retinal pigment epithelium.

[0093] The system may be trained to recognize shadows cast by retinal vessels on the underlying retinal pigment epithelium, which can be more pronounced in cases of retinal detachment due to the altered spatial relationships of retinal structures.

[0094] In some cases, the model may be configured to identify signs of proliferative vitreoretinopathy, such as the presence of fibrovascular proliferation or membrane formation, which can be associated with more advanced or chronic retinal detachment.

[0095] The machine learning system may employ various image processing techniques to enhance its ability to detect these features. For example, contrast enhancement algorithms may be applied to highlight subtle changes in retinal texture or elevation. Edge detection methods may be used to accentuate the boundaries of detached areas or retinal tears.

[0096] In some implementations, the system may utilize a multi-scale approach, analyzing the retinal images at different levels of magnification to capture both large-scale changes (such as overall retinal elevation) and small-scale features (such as microstructural changes in retinal layers).

[0097] The model may be trained on a diverse dataset of retinal images that includes examples of these various features associated with retinal detachment. This training process may enable the system to recognize these indicators across a range of image qualities and patient characteristics.

[0098] In some aspects, the system may provide a detailed analysis of the detected features, including their location, extent, and severity. This information may be presented to clinicians in the form of annotated images or quantitative reports, potentially aiding in the diagnosis and treatment planning process.

[0099] The machine learning model may be designed to continuously leam and improve its ability to detect these features through ongoing training with new; expert-labeledimages. This may allow the system to adapt to variations in imaging techniques and equipment, as well as to recognize newly identified indicators of retinal detachment as they emerge in clinical research.

[0100] In some implementations, the machine learning system may be configured to detect and classify multiple ophthalmic conditions beyond retinal detachment. This expanded capability may enhance the system's utility' in clinical settings, potentially allowing for comprehensive screening of various sight-threatening conditions.

[0101] Central Retinal Artery Occlusion (CRAO)

[0102] The system may be trained to identify signs of CRAO in both retinal fundus images and OCT scans. In fundus images, the model may detect retinal whitening, which can appear as a pale, opaque area in the retina. The system may also be configured to recognize the characteristic cherry -red spot, which occurs due to the contrast between the pale retina and the normal appearance of the fovea.

[0103] The machine learning model may be designed to identify retinal arteriolar attenuation, where the retinal arteries appear narrowed or constricted. Additionally, the system may detect "box-carring," a phenomenon where the blood column in retinal vessels appears segmented.

[0104] In OCT scans, the model may be trained to recognize increased inner retinal hyperreflectivity, which can manifest as brighter-than-normal inner retinal layers. Conversely, the system may identify decreased outer retinal hyperreflectivify, potentially indicating compromised photoreceptor integrity. The model may also detect the presence of inner retinal fluid or neurosensory detachment, which can appear as hyporeflective spaces within or beneath the retina. Hyperreflective foci, which may represent cellular debris or exudates, may also be identified by the system.

[0105] Endophthalmitis

[0106] For the detection of endophthalmitis, the machine learning model may be configured to analyze both anterior and posterior segment features in retinal images. In fundus images, the system may identify the presence of hypopyon, which appears as a layer of white or yellowish material in the anterior chamber. The model may also detect signs of vitritis, manifesting as haze or decreased clarity in the vitreous cavity.

[0107] The system may be trained to recognize retinal hemorrhages associated with endophthalmitis, which can appear as red spots or blotches in the retina. Additionally, themodel may identify signs of retinal vasculitis, such as perivascular sheathing or exudates.

[0108] In OCT scans, the machine learning system may detect hyperreflectivity of the inner retina, which can indicate inflammatory infiltration. The model may also identify vitreous aggregates, appearing as hyperreflective clumps in the vitreous cavity. The system may be configured to recognize the "inverted snowing-cloud" sign, characterized by hyperreflective dots in the vitreous with a distinctive pattern.

[0109] Papilledema

[0110] For the detection of papilledema, the machine learning model may be trained to analyze specific features in both fundus images and OCT scans. In fundus images, the system may identify a swollen and hyperemic (red) optic disc, which can appear elevated and more prominent than normal. The model may be designed to detect the loss of normal disc margin definition, where the boundaries of the optic disc become blurred or indistinct.

[0111] The system may also be configured to recognize the absence or obscuration of venous pulsations, which can be a subtle but important sign of increased intracranial pressure associated with papilledema.

[0112] In OCT scans, the machine learning model may be trained to measure and analyze the thickness of the retinal nerve fiber layer (RNFL), particularly in the peripapillary area. The system may detect abnormal thickening of the RNFL. which is characteristic of papilledema. Additionally, the model may be designed to identify and quantify Bruch's membrane opening (BMO) angulation, which can be altered in cases of papilledema.

[0113] In some implementations, the system may employ advanced image processing techniques to enhance the visibility of subtle features associated with these conditions. For example, contrast enhancement algorithms may be applied to highlight changes in retinal vasculature or optic disc appearance. Edge detection methods may be used to better delineate the boundaries of the optic disc or areas of retinal whitening.

[0114] The machine learning model may utilize a multi-modal approach, combining information from both fundus images and OCT scans to improve diagnostic accuracy. This integrated analysis may allow the system to correlate findings across different imaging modalities, potentially increasing the confidence of its predictions.

[0115] In some cases, the system may provide quantitative measurements of relevant features, such as the degree of retinal thickening in papilledema or the extent of retinal whitening in CRAO. These measurements may be presented alongside visual annotations of the images, offering clinicians a comprehensive view of the detectedabnormalities.

[0116] The model may be trained on a diverse dataset that includes examples of these conditions across various stages of progression and in patients with different demographic characteristics. This approach may enhance the system's ability to detect these conditions across a wide range of clinical presentations.

[0117] In some aspects, the system may incorporate longitudinal analysis capabilities, allowing it to track changes in retinal features over rime. This functionality may be particularly useful for monitoring the progression of conditions like papilledema or assessing the response to treatment in cases of endophthalmitis.

[0118] The machine learning system may be designed with a modular architecture, allowing for the addition of new condition classifiers as they are developed and validated. This flexibility may enable the system to expand its diagnostic capabilities over time, potentially covering an increasingly broad range of ophthalmic conditions.

[0119] Those skilled in the art will recognize that various modifications and variations can be made to this image preparation process without departing from the spirit and scope of the invention. For example, additional steps may be included to address specific imaging artifacts or to optimize the images for particular analysis techniques.

[0120] Machine Learning Model Training

[0121] The process of training a machine learning model for detecting retinal detachment may involve several steps, as illustrated in Figure 2. In some cases, the process may begin with a set of retinal fundus images 210 that serve as input data for training a machine learning model 220.

[0122] The machine learning model 220 may be comprised of various neural network architectures. In some cases, the machine learning model 220 may be a Convolutional Neural Network (CNN). In other cases, the machine learning model 220 may be a ResNet, MobileNet, LeNet, EfficientNet, AmoebaNet, Inception-ResNet, DenseNet, XCeption, InceptionNet, NasNet, SqueezeNet, ShuffleNet, GPipe, GoogleNet, SENet, or WideResNet architecture. The choice of architecture may depend on factors such as the complexity of the retinal detachment detection task, the size of the available training dataset, and computational resources.

[0123] The retinal fundus images 210 used for training may include a diverse set of images representing various stages of retinal detachment, as well as healthy retinas. In some cases, these images may be labeled by expert ophthalmologists to provide ground truth datafor supervised learning.

[0124] The training process may involve feeding the retinal fundus images 210 through the machine learning model 220, which may learn to identify features and patterns associated with retinal detachment. The model may adjust its internal parameters based on the difference between its predictions and the ground truth labels, using techniques such as backpropagation and gradient descent.

[0125] In some cases, the training process may involve multiple iterations or epochs, where the entire dataset is processed multiple times to refine the model’s performance. The training may also incorporate techniques such as data augmentation, where artificial variations of the training images are created to increase the diversity of the training set and improve the model’s generalization capabilities.

[0126] Once the training process is complete, the resulting trained model 230 may be saved for future use. The trained model 230 may represent a set of optimized parameters that can be used to classify new, unseen retinal fundus images for the presence of retinal detachment and retinal tear.

[0127] Those skilled in the art will recognize that various modifications and variations can be made to the training process without departing from the spirit and scope of the invention. For example, different optimization algorithms, learning rate schedules, or regularization techniques may be employed to improve the model’s performance or prevent overfitting.

[0128] Image Analysis and Classification

[0129] The process of analyzing retinal fundus images using the trained machine learning model may involve several steps, as illustrated in Figure 3. In some cases, the process may begin with a trained model 230, which represents the machine learning model that has been previously trained on a dataset of labeled retinal fundus images.

[0130] The trained model 230 may be ported to a computing device, resulting in a ported model 310. This step may involve optimizing the model for the specific hardware and software environment of the target computing device. In some cases, the ported model 310 may be a compressed or quantized version of the original trained model 230 to improve computational efficiency while maintaining accuracy.

[0131] Once the ported model 310 is available on the computing device, the retinal fundus images may be input for analysis. In some cases, these images may be preprocessed using techniques similar to those described in the image preparation process to ensureconsistency with the training data.

[0132] The analysis step 320 may involve feeding the preprocessed retinal fundus images through the ported model 310. During this step, the model may extract relevant features from the images and apply its learned parameters to classify the images. The analysis step 320 may be capable of detecting various types of retinal detachment, including macula- on versus macula-off retinal detachment, rhegmatogenous retinal detachment, and non- rhegmatogenous retinal detachment such as exudative retinal detachment or tractional retinal detachment.

[0133] In addition to detecting various types of retinal detachment, the analysis step 320 may also be capable of identifying other retinal abnormalities that may be precursors to or associated with retinal detachment. These abnormalities may include retinal tears, holes, or breaks.

[0134] Retinal tears may occur when the vitreous gel inside the eye pulls away from the retina, creating a small rip or tear in the retinal tissue. The machine learning model may be trained to recognize the characteristic appearance of retinal tears, which may present as small, irregularly shaped dark areas on the retinal fundus image. In some cases, the model may be able to differentiate between horseshoe tears, which have a flap of tissue still attached to the retina, and operculated tears, where a small piece of retina has become completely detached.

[0135] Retinal holes, which are small, round defects in the retinal tissue, may also be detected by the analysis step 320. These holes may appear as small, well-defined dark spots on the fundus image. The model may be trained to distinguish between atrophic holes, which develop gradually due to thinning of the retina, and traumatic holes, which result from injui ’ to the eye.

[0136] The system may also be capable of identify ing retinal breaks, which are a broader category that includes both tears and holes. In some implementations, the model may classify the severity of retinal breaks based on their size, location, and proximity to blood vessels. This classification may help prioritize cases for further evaluation or treatment.

[0137] In some cases, the analysis step 320 may detect lattice degeneration, a condition characterized by thinning of the peripheral retina that can predispose patients to retinal tears and detachment. The model may be trained to recognize the characteristic appearance of lattice degeneration, which may present as oval or linear areas of retinal thinning with pigmentary changes.

[0138] The system may also be capable of identify ing vitreous traction, where thevitreous gel pulls on the retina without causing a full tear or detachment. This condition may be visible as subtle distortions or elevations of the retinal surface in the fundus image.

[0139] By incorporating the detection of these related retinal abnormalities, the analysis step 320 may provide a more comprehensive assessment of retinal health and potential risk factors for retinal detachment. This expanded capability may enhance the system's utility as a screening tool and aid in early intervention strategies to prevent progression to full retinal detachment.

[0140] The machine learning model may be trained on a diverse dataset that includes examples of these various retinal abnormalities, allowing it to differentiate between different types of lesions and assess their potential risk for progressing to retinal detachment. In some implementations, the model may provide a risk score or classification for each detected abnormality, helping to guide clinical decision-making and follow-up care.

[0141] Following the analysis, an output step 330 may generate the results of the classification. In some cases, the output may include a binary classification (detachment present or absent) along with a confidence score. In other cases, the output may provide more detailed information about the type and extent of the detected retinal detachment.

[0142] In some implementations, the analysis process may involve uploading the retinal fundus images to a cloud network for remote analysis using the machine learning model. This approach may allow for the use of more computationally intensive models or larger datasets that may not be feasible on local computing devices. The cloud-based analysis may also facilitate continuous model updates and improvements based on new data.

[0143] Those skilled in the art will recognize that various modifications and variations can be made to this analysis and classification process without departing from the spirit and scope of the invention. For example, different thresholds for classification confidence may be applied depending on the specific clinical context or requirements.

[0144] Model Updating Process

[0145] The process of updating the machine learning model based on expert analysis may involve several steps, as illustrated in Figure 4. This feedback loop may help improve the model’s accuracy and performance over time.

[0146] In some cases, the process may begin with a retinal fundus image 140, which has been previously analyzed by the machine learning model. The retinal fundus image 140 may then undergo an ophthalmologist analysis 420. During the ophthalmologist analysis 420, an expert ophthalmologist may review the image and provide their professional assessment ofwhether retinal detachment is present, and if so, the type and extent of the detachment.

[0147] The ophthalmologist analysis 420 may serve as a form of ground truth data, which can be compared to the machine learning model’s predictions. In some cases, this comparison may reveal instances where the model’s predictions differ from the expert’s assessment, highlighting areas for potential improvement.

[0148] The process may then incorporate the trained model 230, which represents the cunent version of the machine learning model. Using the insights gained from the ophthalmologist analysis 420, the trained model 230 may be updated to create an updated model 430.

[0149] The updating process may involve various techniques depending on the specific machine learning architecture used. In some cases, the model may undergo finetuning, where the existing model parameters are adjusted slightly based on the new data. In other cases, the model may be retrained on a combination of the original training data and the newly analyzed images.

[0150] The updated model 430 may incorporate the new knowledge gained from the ophthalmologist analysis 420, potentially improving its ability to detect subtle signs of retinal detachment or reducing false positives and false negatives. This iterative process of model updating may allow the system to continuously improve its performance over time as more expert-analyzed images become available.

[0151] In some implementations, the model updating process may be performed periodically, such as weekly or monthly, depending on the volume of new data available and computational resources. In other cases, the updating may occur in real-time as new expert analyses are provided.

[0152] The updated model 430 may then be used for subsequent analyses of retinal fundus images, potentially providing more accurate and reliable detections of retinal detachment. This continuous improvement process may help the system adapt to new patterns or variations in retinal detachment presentations that may not have been present in the original training data.

[0153] Those skilled in the art will recognize that various modifications and variations can be made to this model updating process without departing from the spirit and scope of the invention. For example, different strategies for balancing new and old data in the updating process may be employed to maintain the model’s performance across a wide range of cases.

[0154] The sensitivity and specificity of the machine learning model for detectingretinal detachment may be important metrics for evaluating its performance. In some implementations, the model may achieve a sensitivity and specificity of at least 75% when tested on an independent dataset of retinal fundus images. This lower bound may provide a baseline level of performance that ensures the model’s utility in clinical screening applications. In other implementations, the machine learning model may demonstrate higher levels of accuracy, potentially achieving sensitivity and specificity values of 76%, 77%, 78%, 79%. 80%. 81%. 82%. 83%. 84%. 85%. 86%. 87%. 88%. 89%. 90%. 91%. 92%. 93%. 94%. 95%, 96%, 97%, 98%, 99%, or 100%. These enhanced performance metrics may indicate increasingly robust and reliable detection capabilities, potentially approaching or matching the accuracy of expert human observ ers.

[0155] In other embodiments, the machine learning model may demonstrate higher levels of accuracy, potentially achieving sensitivity and specificity values of 90% or greater. These enhanced performance metrics may indicate a more robust and reliable detection capability, potentially approaching the accuracy of expert human observ ers.

[0156] The specific sensitivity and specificity values may vary depending on factors such as the quality and diversity of the training data, the complexity of the machine learning model architecture, and the characteristics of the test dataset. In some cases, trade-offs between sensitivity and specificity may be made to optimize the model for particular clinical use cases.

[0157] Those skilled in the art may be familiar with methods for calculating sensitivity and specificity. However, for clarity, these metrics may be computed as follows:

[0158] Sensitivity may be calculated using the formula:TruePositivesSensitivity = - - -TruePositives + FalseNegatives

[0159] Specificity may be calculated using the formula:TrueNegativesSpecificity — -TrueNegatives + FalsePositives

[0160] To determine these values, the machine learning model may be evaluated on a test set of labeled retinal fundus images that were not used during the training process. The model’s predictions may be compared to the ground truth labels to identify true positives, true negatives, false positives, and false negatives.

[0161] In some implementations, confidence intervals for sensitivity and specificity may be calculated to provide a range of plausible values given the sample size. Bootstrap resampling or other statistical techniques may be employed to estimate these confidenceintervals.

[0162] The performance of the machine learning model may also be assessed using additional metrics such as the area under the receiver operating characteristic (ROC) curve, which may provide a more comprehensive view of the model’s performance across different classification thresholds.

[0163] It may be important to note that the reported sensitivity’ and specificity values should be based on a sufficiently large and diverse test set to ensure reliable estimates of the model’s performance. In some cases, cross-validation techniques may be employed to provide more robust performance estimates, particularly when working with limited datasets.

[0164] End-to-End Workflow

[0165] The systems and methods described herein may involve an end-to-end workflow for detecting retinal detachment using machine learning, as illustrated in Figure 5. Figure 5 shows a method 500 that encompasses the entire process from image collection to output delivery.

[0166] In some cases, the method 500 may begin with a step 510 of collecting images for training the machine learning model. These images may include a variety of retinal fundus images representing both healthy retinas and those with various stages of retinal detachment.

[0167] The method 500 may continue with a step 520 of cleaning and augmenting the collected data. This step may involve applying image processing techniques to enhance the quality of the retinal fundus images and potentially creating artificial variations to increase the diversity of the training set.

[0168] Following data preparation, a step 530 of training the machine learning model may be performed. During this step, the retinal fundus images may be used to train the machine learning model to recognize patterns and features associated with retinal detachment.

[0169] In some cases, the method 500 may include a step 540 of enforcing learning by calibration. This step may involve fine-tuning the machine learning model’s parameters to optimize its performance on the specific task of retinal detachment detection.

[0170] The method 500 may then proceed to a step 550 of converging the machine learning model. This step may represent the final stage of the training process, where the model’s performance stabilizes and reaches a desired level of accuracy.

[0171] Once the machine learning model is trained and converged, a step 560 ofporting the machine learning system on a computing device may be performed. This step may involve optimizing the model for deployment on specific hardware, such as mobile devices or clinical workstations.

[0172] The method 500 may then move to the operational phase, beginning with a step 570 of inputting a retinal image. This step may involve capturing a new retinal fundus image or retrieving a previously captured image for analysis.

[0173] Following image input, a step 580 of generating classified output using the machine learning system may be performed. During this step, the ported machine learning model may analyze the input retinal fundus image and produce a classification result indicating the presence or absence of retinal detachment.

[0174] In some cases, the method 500 may include a step 585 of seeking ophthalmologist opinion / analysis. This step may involve having an expert ophthalmologist review the machine learning system’s output and provide their professional assessment.

[0175] Based on the ophthalmologist’s analysis, the method 500 may include a step 587 of updating the machine learning system. This step may involve refining the model’s parameters based on any discrepancies between the machine learning output and the expert opinion, potentially improving the system’s accuracy over time.

[0176] The method 500 may conclude with a step 590 of delivering the output to the patient and / or third party. This step may involve communicating the results of the retinal detachment analysis in a clear and understandable format.

[0177] In some cases, the output delivered in step 590 may include recommendations on areas of the ophthalmic image where the retinal condition might be present. These recommendations may help guide further examination or treatment planning. Additionally, the output may include suggestions for positioning the individual to potentially slow dow n the progression of the disease, if retinal detachment is detected.

[0178] The method 500 may end at a step 599, representing the completion of the end-to-end workflow for a single retinal fundus image analysis. However, the process may be repeated for multiple images or as part of ongoing screening programs.

[0179] In some implementations, the results generated in step 590 may be stored in an electronic medical record (EMR) system. This integration with the EMR may allow for longitudinal tracking of retinal health and facilitate comprehensive patient care. The system may automatically update the patient's medical record with the analysis results, potentially including the retinal fundus image, the machine learning model's classification, confidence scores, and any associated recommendations.

[0180] The results may also be transmitted to clinicians remotely for review. This feature may enable telemedicine applications, allowing ophthalmologists or other eye care professionals to assess the machine learning model's output from different locations. The system may incorporate secure data transmission protocols to ensure patient privacy and comply with relevant healthcare regulations.

[0181] In some cases, the system may provide an option to deliver the results in an encrypted format. This encryption may add an extra layer of security, particularly when transmitting sensitive medical information over networks. The encrypted results may be accompanied by a decryption key, which may be securely provided to authorized recipients. This approach may allow for controlled access to the analysis results, ensuring that only intended parties can view the sensitive medical information.

[0182] The encryption method used may vary depending on the specific security requirements and regulatory standards applicable to the healthcare setting. In some implementations, the system may use asymmetric encryption algorithms, where a public key is used to encrypt the results, and a private key is required for decryption. Alternatively, symmetric encryption methods may be employed, with secure key exchange protocols to ensure that the decryption key is safely transmitted to authorized recipients.

[0183] The system may also implement role-based access controls, where different levels of encrypted information are made available to different types of users. For example, a summary of the results may be encrypted with one key for general healthcare providers, while the full detailed analysis may be encrypted with a different key for specialist ophthalmologists.

[0184] In some cases, the encrypted results may include time-limited access tokens, allowing recipients to view the results for a specified period before the decryption key expires. This feature may provide additional control over the dissemination of sensitive medical information and ensure that the most up-to-date analysis is always referenced.

[0185] The integration of EMR storage, remote clinician review, and encrypted result delivery may enhance the versatility and security of the retinal detachment detection system. These features may facilitate efficient information sharing among healthcare providers while maintaining patient confidentiality and data integrity.

[0186] The method 500 may end at a step 599, representing the completion of the end-to-end workflow for a single retinal fundus image analysis. However, the process may be repeated for multiple images or as part of ongoing screening programs.

[0187] Those skilled in the art will recognize that various modifications andvariations can be made to this end-to-end workflow without departing from the spirit and scope of the invention. For example, additional steps may be incorporated to address specific clinical requirements or to integrate with existing healthcare information systems.

[0188] Machine Learning Models

[0189] A machine learning model can comprise a supervised, semi-supervised, unsupervised, or self-supervised machine learning model. In some examples, the machine learning approach comprises a classical machine learning method, such as, but not limited to, support vector machine (SVM) (e.g., one-class SVM, linear or radial kernels, etc.), K-nearest neighbor (KNN). isolation forest, random forest, logistic regression, AdaBoost classifier, extra trees classifier, extreme gradient boosting, gaussian process classifier, gradient boosting classifier, light gradient boosting, linear discriminant analysis, naive Bayes, quadratic discriminant analysis, ridge classifier, or any combination thereof. In some examples, the machine learning approach comprises a deep leaning method (e.g., deep neural network (DNN)), such as. but not limited to a fully-connected network, convolutional neural network (CNN) (e.g.. one-class CNN), recurrent neural network (RNN). transformer, graph neural network (GNN), convolutional graph neural network (CGNN), multi-level perceptron (MLP), or any combination thereof.

[0190] In some embodiments, a classical ML method comprises one or more algorithms that leams from existing observations (i.e., known features) to predict outputs. In some embodiments, the one or more algorithms perform clustering of data. In some examples, the classical ML algorithms for clustering comprise K-means clustering, meanshift clustering, density -based spatial clustering of applications with noise (DBSCAN). expectation-maximization (EM) clustering (e.g., using Gaussian mixture models (GMM)), agglomerative hierarchical clustering, or any combination thereof. In some embodiments, the one or more algorithms perform classification of data. In some examples, the classical ML algorithms for classification comprise logistic regression, naive Bayes, KNN, random forest, isolation forest, decision trees, gradient boosting, support vector machine (SVM), or any combination thereof. In some examples, the SVM comprises a one-class SMV or a multiclass SVM.

[0191] In some embodiments, the deep learning method comprises one or more algorithms that leams by extracting new features to predict outputs. In some embodiments, the deep learning method comprises one or more layers. In some embodiments, the deep learning method comprises a neural network (e.g., DNN comprising more than one layer). Insome examples, the machine learning approach comprises a deep leaning method (e.g.. deep neural network (DNN)), such as, but not limited to a fully-connected network, convolutional neural network (CNN) (e.g., one-class CNN), recunent neural network (RNN), transformer, graph neural network (GNN), convolutional graph neural network (CGNN), multi-level perceptron (MLP), or any combination thereof. Neural networks generally comprise connected nodes in a network, which can perform functions such as transforming or translating input data. In some embodiments, the output from a given node is passed on as input to another node. The nodes in the network generally comprise input units in an input layer, hidden units in one or more hidden layers, output units in an output layer, or a combination thereof. In some embodiments, an input node is connected to one or more hidden units. In some embodiments, one or more hidden units is connected to an output unit. The nodes can generally take in input through the input units and generate an output from the output units using an activation function. In some embodiments, the input or output comprises a tensor, a matrix, a vector, an array, or a scalar. In some embodiments, the activation function is a Rectified Linear Unit (ReLU) activation function, a sigmoid activation function, a hyperbolic tangent activation function, or a Softmax activation function. In some embodiments, the deep learning methods can include a Vision Transformer (ViT).

[0192] The connections between nodes can further comprise weights for adjusting input data to a given node (i.e., to activate input data or deactivate input data). In some embodiments, the weights are learned by the neural network. In some embodiments, the neural network is trained to learn weights using gradient-based optimizations. In some embodiments, the gradient-based optimization comprises one or more loss functions. In some embodiments, the gradient-based optimization is gradient descent, conjugate gradient descent, stochastic gradient descent, or any variation thereof (e.g., adaptive moment estimation (Adam)). In some further embodiments, the gradient in the gradient-based optimization is computed using backpropagation. In some embodiments, the nodes are organized into graphs to generate a network (e.g., graph neural networks). In some embodiments, the nodes are organized into one or more layers to generate a network (e.g., feed forw ard neural networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), etc.). In some embodiments, the CNN comprises a one-class CNN or a multi-class CNN.

[0193] In some embodiments, the neural network comprises one or more recurrent layers. In some embodiments, the one or more recurrent layers are one or more long short-term memory (LSTM) layers or gated recurrent units (GRUs). In some embodiments, the one or more recurrent layers perform sequential data classification and clustering in which the data ordering is considered (e.g., time series data). In such embodiments, future predictions are made by the one or more recurrent layers according to the sequence of past events. In some embodiments, the recurrent layer retains or “remembers” important information, while selectively “forgets” what is not essential to the classification.

[0194] In some embodiments, the neural network comprises one or more convolutional layers. In some embodiments, the input and the output are a tensor representing variables or attributes in a data set (e.g., features), which may be referred to as a feature map (or activation map). In such embodiments, the one or more convolutional layers are referred to as a feature extraction phase. In some embodiments, the convolutions are one-dimensional (ID) convolutions, two dimensional (2D) convolutions, three dimensional (3D) convolutions, or any combination thereof. In further embodiments, the convolutions are ID transpose convolutions, 2D transpose convolutions, 3D transpose convolutions, or any combination thereof.

[0195] The layers in a neural network can further comprise one or more pooling layers before or after a convolutional layer. In some embodiments, the one or more pooling layers reduces the dimensionality of a feature map using filters that summarize regions of a matrix. In some embodiments, this down samples the number of outputs, and thus reduces the parameters and computational resources needed for the neural network. In some embodiments, the one or more pooling layers comprises max pooling, min pooling, average pooling, global pooling, norm pooling, or a combination thereof. In some embodiments, max pooling reduces the dimensionality of the data by taking only the maximums values in the region of the matrix. In some embodiments, this helps capture the most significant one or more features. In some embodiments, the one or more pooling layers is one dimensional (ID), two dimensional (2D), three dimensional (3D), or any combination thereof.

[0196] The neural network can further comprise of one or more flattening layers, which can flatten the input to be passed on to the next layer. In some embodiments, an input (e.g.. feature map) is flattened by reducing the input to a one-dimensional array. In some embodiments, the flattened inputs can be used to output a classification of an object. In some embodiments, the classification comprises a binary classification or multi-class classification of visual data (e.g.. images, videos, etc.) or non-visual data (e.g., analogue sensor measurements, etc.). In some embodiments, the classification comprises binary classification of an image (e.g., cat or dog). In some embodiments, the classification comprises binaryclassification of a measurement. In some examples, the binary classification of a measurement comprises a classification of a system's performance using the physical measurements described herein (e.g., normal or abnormal).

[0197] The neural networks can further comprise of one or more dropout layers. In some embodiments, the dropout layers are used during training of the neural network (e.g., to perform binary or multi-class classifications). In some embodiments, the one or more dropout layers randomly set some weights as 0 (e.g.. about 10%, 20%, 30%, 40%, 50%, 60%, 70%, or 80% of weights). In some embodiments, the setting some weights as 0 also sets the corresponding elements in the feature map as 0. In some embodiments, the one or more dropout layers can be used to avoid the neural network from overfitting.

[0198] The neural network can further comprise one or more dense layers, which comprises a fully connected network. In some embodiments, information is passed through a fully connected network to generate a predicted classification of an object. In some embodiments, the error associated with the predicted classification of the object is also calculated. In some embodiments, the error is backpropagated to improve the prediction. In some embodiments, the one or more dense layers comprises a Softmax activation function. In some embodiments, the Softmax activation function converts a vector of numbers to a vector of probabilities. In some embodiments, these probabilities are subsequently used in classifications, such as classifications of the retinal detachment analysis as described herein.

[0199] The machine learning model can comprise one or more sub-models. In some cases, the one or more sub-models are trained individually. In some cases, individual models are trained to determine whether a retina has been detached, or may have an anomaly as described herein. In some embodiments, a single model is trained to directly analyze retinal detachment in a single model.

[0200] Further optional features could include the use of transformer-based natural language processing and conversational Al, such as systems including large language models including ChatGPT, Perplexity, Grok, and the like. Virtual Reality (object removal) systems could include YOLO and 3D-aware inpainting systems known to those of skill in the art.

[0201] Apparatus for Image Capture and Analysis

[0202] The systems and methods described herein may include an apparatus for capturing and analyzing retinal fundus images, as illustrated in Figure 6. This apparatus may provide a portable and efficient means of obtaining high-quality retinal images for subsequent analysis using the machine learning techniques described earlier.

[0203] In some cases, the apparatus may include a patient indicator 600, which may represent the position where an individual's eye is placed for imaging. The patient indicator 600 may be designed to comfortably support the individual’s head and align their eye with the imaging components of the apparatus.

[0204] The apparatus may include a magnification lens 610, which may be a high- power lens configured to view the individual’s retina. In some cases, the magnification lens 610 may have a magnification power ranging from 20D to 30D, allowing for detailed visualization of retinal structures.

[0205] A lens slider 620 may be provided to hold and adjust the position of the magnification lens 610. The lens slider 620 may be designed to move along a shaft 630, allowing for precise positioning of the magnification lens 610 relative to the individual’s eye.

[0206] The apparatus may also include a device holder 640, which may be configured to secure an electronic device 660. In some cases, the electronic device 660 may be a smartphone, tablet, or other portable computing device equipped with a camera. The device holder 640 may be designed to slide along the shaft 630. enabling adjustment of the electronic device 660’ s position relative to the magnification lens 610.

[0207] An imaging camera 670 may be part of the electronic device 660. The imaging camera 670 may be used to capture retinal fundus images through the magnification lens 610. The ability to adjust both the lens slider 620 and the device holder 640 along the shaft 630 may allow for precise alignment of the imaging camera 670 with the magnification lens 61 , facilitating the capture of clear and detailed retinal images.

[0208] In some cases, an analysis application 650 may run on the electronic device 660. The analysis application 650 may be designed to control the image capture process, process the captured retinal fundus images, and potentially perform initial analysis using the machine learning techniques described earlier.

[0209] The apparatus may be designed to allow for obtaining retinal fundus images by capturing retinal images with the electronic device 660 comprising the imaging camera 670, in conjunction with the portable components including the magnification lens 610 and lens slider 620. The adjustability of the device holder 640 and lens slider 620 along the shaft 630 may allow for precise alignment of the imaging component (magnification lens 610) with the imaging camera 670.

[0210] In some cases, the analysis application 650 may incorporate the trained model 230 or ported model 310 described earlier, allowing for on-device analysis of the captured retinal fundus images. This integration may enable rapid screening for retinaldetachment or other retinal conditions at the point of care.

[0211] The portability and adjustability of this apparatus may facilitate its use in various clinical settings, including remote or resource-limited environments where access to traditional ophthalmic imaging equipment may be limited. This may expand the reach of retinal detachment screening and potentially improve early detection rates.

[0212] Those skilled in the art will recognize that various modifications and variations can be made to this apparatus without departing from the spirit and scope of the invention. For example, different lens magnifications or camera specifications may be used to optimize image quality for specific clinical needs or to accommodate advancements in mobile device technology. For example, the systems and methods described herein may be compatible with various existing retinal imaging devices and techniques. In some implementations, the apparatus for capturing retinal fundus images may be used in conjunction with or as an alternative to traditional retinal fundus cameras. These conventional fundus cameras may include tabletop models with chin rests and adjustable optics, as well as handheld portable versions designed for increased mobility.

[0213] In addition to fundus photography, the system may incorporate or interface with optical coherence tomography (OCT) devices. OCT may provide high-resolution, cross- sectional images of the retina, allowing for detailed analysis of retinal layers and structures. The integration of OCT data with fundus images may enhance the system's ability to detect and characterize retinal detachment and other retinal pathologies.

[0214] The system may also be compatible with ultra-widefield (UWF) imaging technologies, which can capture a broader view of the retina in a single image. UWF imaging may be particularly useful for detecting peripheral retinal detachments that might be missed by standard fundus photography.

[0215] In some cases, the system may incorporate or interface with fluorescein angiography equipment. This technique may involve injecting a fluorescent dye into the bloodstream and capturing a series of retinal images to visualize blood flow' and identify areas of retinal detachment or other vascular abnormalities.

[0216] The apparatus and methods described herein may also be adapted for use with scanning laser ophthalmoscopy (SLO) devices. SLO may provide high-contrast images of the retina and may be particularly useful for detecting subtle changes in retinal structure.

[0217] In some implementations, the system may be designed to work with adaptive optics retinal imaging systems. These advanced devices may use deformable mirrors to correct for optical aberrations in the eye, potentially allowing for even higher-resolutionimaging of retinal structures.

[0218] The machine learning models and analysis techniques described in this disclosure may be applied to images obtained from any of these imaging modalities, either individually or in combination. The system may be designed with modular components to allow for easy integration with various imaging devices and technologies.

[0219] Those skilled in the art will recognize that there are numerous other retinal imaging techniques and devices that may be compatible with or incorporated into the systems and methods described herein. The specific choice of imaging modality may depend on factors such as the clinical setting, the specific retinal conditions being assessed, and the available resources. The system may be designed to be flexible and adaptable to accommodate advancements in retinal imaging technology as they emerge.

[0220] Computer systems

[0221] The present disclosure provides computer systems that are programmed to implement methods of the disclosure. Figure 7 shows a computer system 701 that is programmed or otherwise configured to perform segmentation and / or classification of one or more inputs, and / or to provide reports and functionality' to end users as described herein. The computer system 701 can regulate various aspects of the systems described herein by methods and systems of the present disclosure. The computer system 701 can be an electronic device of a user or a computer system that is remotely located with respect to the electronic device. The electronic device can be a mobile electronic device.

[0222] The computer system 701 includes a central processing unit (CPU, also “processor’ and “computer processor” herein) 705, which can be a single core or multi core processor, or a plurality of processors for parallel processing. The computer system 701 also includes memory' or memory location 710 (e.g., random-access memory, read-only memory, flash memory), electronic storage unit 715 (e.g., hard disk), communication interface 720 (e.g., network adapter) for communicating with one or more other systems, and peripheral devices 725, such as cache, other memory, data storage and / or electronic display adapters. The memory 710, storage unit 715, interface 720 and peripheral devices 725 are in communication with the CPU 705 through a communication bus (solid lines), such as a motherboard. The storage unit 715 can be a data storage unit (or data repository') for storing data. The computer system 701 can be operatively coupled to a computer network (“network”) 730 with the aid of the communication interface 720. The network 730 can be the Internet, an internet and / or extranet, or an intranet and / or extranet that is incommunication with the Internet. The network 730 in some cases is a telecommunication and / or data network. The network 730 can include one or more computer servers, which can enable distributed computing, such as cloud computing. The network 730, in some cases with the aid of the computer system 701, can implement a peer-to-peer network, which may enable devices coupled to the computer system 701 to behave as a client or a server.

[0223] The CPU 705 can execute a sequence of machine-readable instructions, which can be embodied in a program or software. The instructions may be stored in a memory location, such as the memory 710. The instructions can be directed to the CPU 705, which can subsequently program or otherwise configure the CPU 705 to implement methods of the present disclosure. Examples of operations performed by the CPU 705 can include fetch, decode, execute, and writeback.

[0224] The CPU 705 can be part of a circuit, such as an integrated circuit. One or more other components of the system 701 can be included in the circuit. In some cases, the circuit is an application specific integrated circuit (ASIC).

[0225] A graphics processing unit (GPU) is a specialized processing unit, electronic circuit, module, or computer chip, etc., that can accelerate digital image processing and many other applications and is often present either as a discrete video card, or embedded on motherboards, or as integrated graphics on a CPU. Similarly, chip modules are known that can perform machine learning prediction (sometimes referred to as inference). Such chips include, for example, language processing units (LPUs), cloud tensor processing units (TPUs), neural engines, Al coprocessors, Al accelerators, and neural processing units (NPUs). In some embodiments, a GPU or other chip module performs at least some of the functions that could otherwise be performed by a CPU.

[0226] The storage unit 715 can store files, such as drivers, libraries and saved programs. The storage unit 715 can store user data, e.g., user preferences and user programs. The computer system 701 in some cases can include one or more additional data storage units that are external to the computer system 701, such as located on a remote server that is in communication with the computer system 701 through an intranet or the Internet.

[0227] The computer system 701 can communicate with one or more remote computer systems through the network 730. For instance, the computer system 701 can communicate with a remote computer system of a user (e.g., a lab technician or a treating physician or opthalmologist). Examples of remote computer systems include personal computers (e.g., portable PC), slate or tablet PC's (e.g., Apple® iPad, Samsung® Galaxy Tab), telephones. Smart phones (e.g., Apple® iPhone, Android-enabled device.Blackberry®), or personal digital assistants. The user can access the computer system 101 via the network 730.

[0228] Methods as described herein can be implemented by way of machine (e.g., computer processor) executable code stored on an electronic storage location of the computer system 701, such as, for example, on the memory 710 or electronic storage unit 715. The machine executable or machine-readable code can be provided in the form of software. During use, the code can be executed by the processor 705. In some cases, the code can be retrieved from the storage unit 715 and stored on the memory 710 for ready access by the processor 705. In some situations, the electronic storage unit 715 can be precluded, and machine-executable instructions are stored on memory' 710.

[0229] The code can be pre-compiled and configured for use with a machine having a processer adapted to execute the code or can be compiled during runtime. The code can be supplied in a programming language that can be selected to enable the code to execute in a pre-compiled or as-compiled fashion.

[0230] Aspects of the systems and methods provided herein, such as the computer system 701, can be embodied in programming. Various aspects of the technology may be thought of as “products” or “articles of manufacture” typically in the form of machine (or processor) executable code and / or associated data that is carried on or embodied in a ty pe of machine readable medium. Machine-executable code can be stored on an electronic storage unit, such as memory’ (e.g., read-only memory, random-access memory, flash memory) or a hard disk. “Storage” type media can include any or all of the tangible memory of the computers, processors or the like, or associated modules thereof, such as various semiconductor memories, tape drives, disk drives and the like, which may provide non- transitory storage at any time for the software programming. All or portions of the software may at times be communicated through the Internet or various other telecommunication networks. Such communications, for example, may enable loading of the software from one computer or processor into another, for example, from a management server or host computer into the computer platform of an application server. Thus, another type of media that may bear the software elements includes optical, electrical and electromagnetic waves, such as used across physical interfaces between local devices, through wired and optical landline networks and over various air-links. The physical elements that carry' such waves, such as wired or wireless links, optical links or the like, also may be considered as media bearing the software. As used herein, unless restricted to non-transitory, tangible “storage” media, terms such as computer or machine “readable medium” refer to any medium that participates inproviding instructions to a processor for execution.

[0231] Hence, a machine readable medium, such as computer-executable code, may take many forms, including but not limited to, a tangible storage medium, a carrier wave medium, or physical transmission medium. Non-volatile storage media include, for example, optical or magnetic disks, such as any of the storage devices in any computer(s) or the like, such as may be used to implement the databases, etc. shown in the drawings. Volatile storage media include dynamic memory, such as main memory of such a computer platform. Tangible transmission media include coaxial cables, copper wire and fiber optics, including the wires that comprise a bus within a computer system. Carrier-wave transmission media may take the form of electric or electromagnetic signals, or acoustic or light waves such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media therefore include for example: a floppy disk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a CD-ROM, DVD or DVD-ROM, any other optical medium, punch cards paper tape, any other physical storage medium with patterns of holes, a RAM, a ROM, a PROM and EPROM, a FLASH-EPROM, any other memory chip or cartridge, a carrier wave transporting data or instructions, cables or links transporting such a carrier wave, or any other medium from which a computer may read programming code and / or data. Many of these forms of computer readable media may be involved in carrying one or more sequences of one or more instructions to a processor for execution.

[0232] The computer system 701 can include or be in communication with an electronic display that comprises a user interface (UI) 740 for providing, for example, evaluation of retina data, segmentation of such an image, and / or automatic determination of data associated with the retina detachment. Examples of UTs include, without limitation, a graphical user interface (GUI) and w eb-based user interface.

[0233] Methods and systems of the present disclosure can be implemented by way of one or more algorithms. An algorithm can be implemented by w ay of software upon execution by the central processing unit 705. The algorithm can, for example, analyze retinal properties of a subject sample using the methods provided herein.

[0234] System Integration and Function

[0235] The systems and methods described herein may integrate various hardware and software components to create a comprehensive solution for detecting retinal detachment using machine learning techniques. This integration may allow for a synergistic approach thatcombines the advantages of advanced imaging technology with sophisticated data analysis capabilities.

[0236] In some cases, the hardware components, such as the retinal imaging apparatus, may work in conjunction with the machine learning software to capture high- quality retinal fundus images and analyze them for signs of retinal detachment. The adjustable nature of the imaging apparatus may allow for precise alignment and focus, potentially improving the quality of the input data for the machine learning algorithms.

[0237] The machine learning software may be designed to process the captured retinal fundus images, extracting relevant features and applying trained models to detect potential retinal detachment. This software may be integrated with the imaging hardware through dedicated applications running on the electronic device used for image capture.

[0238] In some implementations, the system may be configured to process a stream of retinal fundus images for analysis. This approach may allow for continuous monitoring of retinal health or enable high-throughput screening in clinical settings. The stream of images may be acquired from a single patient over time or from multiple patients in succession.

[0239] When processing a stream of images, the system may employ various techniques to handle the increased data volume and maintain efficiency. In some cases, the system may utilize a sliding window approach, where a fixed number of consecutive images are analyzed together. This method may help in detecting subtle changes or anomalies that might not be apparent in a single image.

[0240] The system may also implement real-time processing capabilities, where each image in the stream is analyzed as soon as it is acquired. This approach may be particularly useful in scenarios where immediate feedback is required, such as during surgical procedures or emergency assessments.

[0241] In some implementations, the system may be designed to handle combined images from multiple patients. This may occur in situations where batch processing is more efficient or when analyzing population-level trends. When dealing with multi-patient image sets, the system may employ several strategies for image separation and analysis.

[0242] One approach may involve separating the images by patient before analysis. In this case, the system may use metadata associated with each image, such as patient identifiers or timestamps, to group images belonging to the same individual. Once grouped, each patient's image set may be processed independently through the machine learning model.

[0243] Alternatively, the system may perform analysis on the combined image setand separate the results by patient afterwards. This method may be useful when the machine learning model is designed to detect general features of retinal detachment rather than patient-specific characteristics. After analysis, the system may associate the results with individual patients using image metadata or other identifying information.

[0244] In some cases, the system may employ a hybrid approach, where initial feature extraction is performed on the combined image set, but final classification or diagnosis is done on a per-patient basis. This method may leverage the efficiency of batch processing while still allowing for personalized analysis.

[0245] The system may also incorporate techniques to handle potential variations in image qualify or characteristics across different patients or imaging sessions. This may include normalization procedures to standardize image features or the use of robust machine learning models that can adapt to diverse input data.

[0246] When processing multi-patient image sets, the system may implement additional privacy and security measures. This may include anonymization techniques to remove personally identifiable information from images before batch processing, or encryption methods to ensure that patient data remains protected throughout the analysis pipeline.

[0247] In some implementations, the system may use the multi-patient image analysis to generate aggregate statistics or identify population-level trends in retinal health. This information may be valuable for epidemiological studies or for assessing the effectiveness of screening programs.

[0248] The ability to handle image streams and multi-patient datasets may enhance the versatility' and efficiency of the retinal detachment detection system. These features may enable the system to adapt to various clinical workflows and research applications, potentially improving the speed and scale of retinal health assessments.

[0249] In some cases, the system may incorporate cloud-based processing capabilities, allowing for more computationally intensive analysis to be performed remotely. This integration of local hardware with cloud-based software may enable the use of more complex machine learning models while maintaining the portability’ and ease of use of the imaging apparatus.

[0250] The determination of the retinal condition may be provided through various interfaces, enhancing the flexibility and utility of the system. In some cases, the results may be displayed on a web interface, allowing for remote access and consultation by healthcare professionals. Alternatively, the determination may be presented on a mobile interface.potentially enabling on-the-go screening and diagnosis.

[0251] In some cases, the electronic device used for image capture may also serve as the platform for displaying the analysis results. This integration of imaging and result presentation on a single device may streamline the workflow and improve efficiency in clinical settings.

[0252] The system may also be designed to communicate the determination of the retinal condition via medical personnel. This integration of automated analysis with professional medical interpretation may provide a comprehensive approach to retinal detachment detection, potentially improving accuracy and patient care.

[0253] The synergy' between the hardware components and machine learning software may extend to the continuous improvement of the system. As new retinal fundus images are captured and analyzed, the results may be used to update and refine the machine learning models, potentially improving the system’s performance over time.

[0254] In some cases, the integrated system may incorporate safeguards to ensure the quality and reliability’ of the retinal detachment detection process. These may include automated checks for image quality, alerts for potentially unreliable results, and mechanisms for seeking expert review in ambiguous cases.

[0255] The integration of these various components may create a system that is potentially more effective and efficient than traditional methods of retinal detachment detection. By combining advanced imaging technology with sophisticated machine learning algorithms and flexible result presentation options, the system may offer a comprehensive solution for early detection and management of retinal detachment.

[0256] The system may be configured to receive images for analysis from a plurality of media sources, enhancing its flexibility and adaptability in various clinical and research settings. In some implementations, the system may accept retinal fundus images from cloud storage platforms, allowing for seamless integration with existing healthcare information systems and facilitating remote image analysis.

[0257] Local storage on the electronic device or a connected computer system may serve as another source for input images. This approach may be particularly useful in settings where internet connectivity is limited or when working with sensitive patient data that must remain on-site.

[0258] The system may also be designed to interface with removable storage devices such as USB drives, SD cards, or external hard drives. This capability may enable easy transfer of images from standalone retinal imaging devices that are not directlyconnected to the analysis system.

[0259] In some cases, the system may receive images through wireless transmission protocols. This may include Wi-Fi networks, Bluetooth connections, or cellular data networks, potentially allowing for real-time image transfer from remote locations or mobile screening units.

[0260] The system may also incorporate the ability’ to receive images through wired connections. This may involve direct connections to retinal imaging devices using standard interfaces such as USB, HDMI, or proprietary connectors. Wired connections may provide high-speed, reliable image transfer in clinical settings where multiple imaging devices are used.

[0261] In some implementations, the system may be capable of receiving images through a combination of these media sources. For example, it may simultaneously accept images from local storage, cloud platforms, and wireless transfers, allowing for flexible workflow integration in diverse healthcare environments.

[0262] The system may include intelligent routing capabilities to manage image inputs from multiple sources. This may involve prioritizing certain input streams based on predefined criteria, such as image quality, urgency of analysis, or source reliability.

[0263] To accommodate various image formats and metadata structures from different sources, the system may incorporate robust image processing pipelines. These pipelines may include format conversion, metadata extraction, and quality assessment steps to ensure that images from all sources can be effectively analyzed by the machine learning models.

[0264] The system may also implement security measures tailored to each input method. For cloud and wireless transfers, this may’ include encryption protocols and secure authentication mechanisms. For local and removable storage, the system may employ access controls and data integrity checks.

[0265] In some cases, the system may maintain a log of image sources and transfer methods for each analyzed image. This information may be useful for auditing purposes, troubleshooting, and optimizing the system's performance across different input modalities.

[0266] The ability to receive images from diverse media sources may enhance the system's utility’ in various scenarios, from large hospital networks with centralized image repositories to mobile screening programs in remote areas. This flexibility’ may contribute to the system's potential for improving access to retinal detachment screening and diagnosis across a wide range of healthcare settings.

[0267] Those skilled in the art will recognize that various modifications and variations can be made to this system integration and function without departing from the spirit and scope of the invention. For example, different hardware configurations or software architectures may be employed to optimize performance in specific clinical settings or to accommodate advancements in technology'.

[0268] EXAMPLES

[0269] Aspects of the present teachings may be further understood in light of the following examples, which should not be construed as limiting the scope of the present teachings in any way.

[0270] EXAMPLE 1: Training and Evaluating a Machine Learning Model for Retinal Detachment Detection

[0271] A convolutional neural network (CNN) is trained to detect retinal detachment from fundus images. A dataset of 10,000 labeled retinal fundus images is collected, with 5,000 images showing retinal detachment and 5,000 showing healthy retinas. The images are preprocessed by resizing to 224x224 pixels and normalizing pixel values.

[0272] A ResNet-50 architecture pretrained on ImageNet is used as the base model. The final fully connected layer is replaced with a new layer with 2 output nodes (detachment / no detachment). The model is fine-tuned on the retinal image dataset using stochastic gradient descent with a learning rate of 0.001 for 50 epochs.

[0273] Data augmentation techniques, including random rotations, flips, and brightness adjustments, may be applied to increase the diversity of the training set and improve the model's generalization capabilities. The dataset is split into training (80%), validation (10%), and test (10%) sets.

[0274] During training, the model’s performance is monitored on the validation set to prevent overfitting. Early stopping may be implemented if the validation loss does not improve for a specified number of epochs. The model with the best validation performance is saved for final evaluation.

[0275] After training, the model is evaluated on the held-out test set of 1,000 labeled fundus images (500 with detachment, 500 without). The trained CNN achieves 92% sensitivity and 94% specificity on this test set.

[0276] To interpret the model’s decisions, Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations are generated. These heatmaps highlight the regions ofthe input images that are most influential in the model’s predictions, which may correspond to areas of retinal detachment.

[0277] The trained model is then deployed in a clinical setting for prospective evaluation. A new set of 200 retinal fundus images, previously unseen by the model, is collected from patients undergoing routine eye examinations. These images are preprocessed using the same pipeline as the training data and then analyzed by the deployed model.

[0278] For each image, the model outputs a probability score indicating the likelihood of retinal detachment. A threshold of 0.5 is initially set for binary classification, where scores above 0.5 are classified as detachment and scores below 0.5 as no detachment.

[0279] The model’s predictions on this prospective dataset are compared to diagnoses made by two independent ophthalmologists who are blinded to the model’s output. In cases where the ophthalmologists disagree, a third expert is consulted to reach a consensus diagnosis.

[0280] The model demonstrates a sensitivity of 90% and specificity of 93% on this prospective dataset, closely matching its performance on the original test set. The model correctly identifies 45 out of 50 cases of retinal detachment, including 3 cases that were initially missed by one of the ophthalmologists but confirmed upon further review.

[0281] In 10 cases, the model’s probability scores are between 0.4 and 0.6, indicating lower confidence. These cases are flagged for more detailed review by ophthalmologists, potentially improving the overall accuracy of the screening process.

[0282] The Grad-CAM visualizations for true positive cases show consistent activation in areas corresponding to retinal detachment, as confirmed by the ophthalmologists. This provides additional confidence in the model’s decision-making process and may assist clinicians in localizing areas of concern.

[0283] Based on these results, the model may be integrated into the clinical workflow as a screening tool, with all positive predictions and low-confidence cases being referred for expert review. The model's performance continues to be monitored, and it may be periodically retrained with newly acquired, expert-labeled data to maintain and potentially improve its accuracy over time.

[0284] EXAMPLE 2: Portable Retinal Imaging Device Performance Assessment

[0285] The portable retinal imaging device described in Figure 6 is tested for image quality and ease of use. Twenty volunteers with no prior ophthalmic imaging experience arerecruited to capture retinal images of 10 test subjects (5 with known retinal detachment, 5 without) using the device.

[0286] Each volunteer captures 3 images of each test subject’s right eye. Image quality is assessed by tw o expert ophthalmologists on a 5-point scale (1 = unusable, 5 = excellent quality). The average image quality7score across all captured images is expected to be 4.2, with 85% of images scoring 3 or higher.

[0287] The time taken to capture each image set is recorded. The average time per image set is expected to decrease from 5 minutes for the first test subject to 2 minutes for the tenth, demonstrating a rapid learning curve for device operation.

[0288] The retinal images captured by the portable device are then analyzed using the CNN from Example 1. The model’s performance on these images is compared to its performance on images from standard clinical fundus cameras. The sensitivity and specificity on the portable device images are expected to be within 5 percentage points of the results on clinical-grade images.

[0289] EXAMPLE 3: End-to-End System Workflow and Performance in a Clinical Setting

[0290] The complete retinal detachment detection system, including the portable imaging device and CNN-based analysis softw are, is deployed in a community health center for a 3-month trial. The center serves a population with a high prevalence of diabetes, a risk factor for retinal detachment.

[0291] During the trial, 500 patients undergoing routine eye exams have retinal images captured using both the portable device and a standard fundus camera. The images from both devices are analyzed by the CNN model and by two expert ophthalmologists who are blinded to the imaging method and model predictions.

[0292] The workflow' efficiency is measured by recording the time taken for image capture, CNN analysis, and result delivery. The average time from patient positioning to result delivery is expected to be under 5 minutes for the portable device system, compared to 15 minutes for the standard workflow.

[0293] The CNN’s performance is evaluated against the consensus diagnosis of the tw o ophthalmologists. On images from the portable device, the CNN is expected to achieve a sensitivity of 91% and specificity7of 93% for detecting retinal detachment. On standard fundus camera images, the sensitivity and specificity are expected to be 93% and 95% respectively.

[0294] Importantly, the CNN is expected to flag 98% of cases that require immediate intervention (e.g., macula-on retinal detachments) for urgent ophthalmologist review. This demonstrates the system’s potential to prioritize high-risk cases in resourcelimited settings.

[0295] A survey of healthcare providers using the system reveals an average satisfaction score of 4.5 out of 5, with particular appreciation for the rapid results and the system’s ability to triage urgent cases. Patient satisfaction scores for the imaging experience are expected to be comparable between the portable device and standard fundus camera.

[0296] EXAMPLE 4: Integration with Existing Retinal Imaging Infrastructure

[0297] The retinal detachment detection system is integrated with the existing retinal imaging infrastructure at a large metropolitan hospital network. This network comprises 5 hospitals and 1 outpatient clinics, each equipped with various retinal imaging devices including standard fundus cameras, ultra-widefield imaging systems, and optical coherence tomography (OCT) machines.

[0298] A cloud-based integration platform is developed to interface with the hospital network's existing picture archiving and communication system (PACS). This platform is designed to receive images from all connected retinal imaging devices, process them through the CNN model, and return results to the ordering physician's workstation.

[0299] Over a 6-month period, the integrated system processes 10,000 retinal images from various sources. The images are automatically routed to the cloud platform, where they undergo preprocessing to standardize format and resolution before analysis by the CNN model.

[0300] The system's performance is evaluated in terms of integration success, processing speed, and diagnostic accuracy. Integration success is measured by the percentage of images successfully transferred, processed, and returned to the originating workstation. The system is expected to achieve a 99.5% success rate in this regard.

[0301] Processing speed is assessed by measuring the time from image acquisition to result delivery. For on-premise processing (using edge computing devices installed at each location), the average turnaround time is expected to be under 60 seconds. For cloud-based processing, the average turnaround time is expected to be under 2 minutes, accounting for secure data transfer times.

[0302] Diagnostic accuracy is evaluated by comparing the CNN model's predictions to the final diagnoses made by ophthalmologists. The model is expected to maintain asensitivity of 92% and specificity of 94% across the diverse range of imaging devices and patient populations in the hospital network.

[0303] The system's ability to handle various image types is tested by processing a subset of 1,000 images that includes standard fundus photographs, ultra-wi defield images, and OCT scans. The CNN model, initially trained on standard fundus images, is fine-tuned using transfer learning techniques to adapt to these different modalities. After fine-tuning, the model is expected to achieve comparable performance across all image types, with sensitivity and specificity varying by no more than 3 percentage points between modalities.

[0304] A key feature of the integration is the system's ability to prioritize urgent cases. The CNN model flags suspected cases of acute retinal detachment, which are then fast- tracked for immediate review by on-call ophthalmologists. This triage system is expected to reduce the average time to treatment for urgent cases by 45 minutes compared to the standard workflow.

[0305] Healthcare provider satisfaction with the integrated system is assessed through surveys and interviews. The system receives an average satisfaction score of 4.3 out of 5, with particular praise for its seamless integration with existing workflows and its ability to prioritize urgent cases. Some providers express initial skepticism about relying on AI- assisted diagnosis, highlighting the need for ongoing education and transparency about the system's capabilities and limitations.

[0306] The successful integration of the retinal detachment detection system with the hospital network's existing infrastructure demonstrates the potential for Al-assisted diagnosis to enhance efficiency and potentially improve patient outcomes in large-scale healthcare settings. The cloud-based approach allows for centralized model updates and performance monitoring, while the option for on-premise processing addresses data privacy concerns and enables rapid analysis in time-sensitive situations.

[0307] Other Embodiments

[0308] The detailed description set-forth above is provided to aid those skilled in the art in practicing the present invention. However, the invention described and claimed herein is not to be limited in scope by the specific embodiments herein disclosed because these embodiments are intended as illustration of several aspects of the invention. Any equivalent embodiments are intended to be within the scope of this invention. Indeed, various modifications of the invention in addition to those shown and described herein will become apparent to those skilled in the art from the foregoing description which do not depart fromthe spirit or scope of the present inventive discovery'. Such modifications are also intended to fall within the scope of the appended claims.

[0309] References Cited

[0310] All publications, patents, patent applications and other references cited in this application are incorporated herein by reference in their entirety for all purposes to the same extent as if each individual publication, patent, patent application or other reference was specifically and individually indicated to be incorporated by reference in its entirety for all purposes. Citation of a reference herein shall not be construed as an admission that such is prior art to the present invention.

Claims

CLAIMSWhat is claimed is:

1. A system for detecting ophthalmic conditions, comprising: an imaging device configured to capture a retinal fundus image of an individual; a processor; and a memory storing instructions that, when executed by the processor, cause the system to: preprocess the retinal fundus image to enhance retinal features; input the preprocessed retinal fundus image into a trained machine learning model; generate, using the trained machine learning model, a determination of whether the retinal fundus image indicates presence of retinal detachment; and output the determination to a user interface.

2. The system of claim 1, wherein the ophthalmic condition is selected from the group consisting of retinal detachment, Central Retinal Artery Occlusion (CRAO), Endophthalmitis, and Papilledema.

3. The system of claims 1 or 2, wherein the imaging device comprises a portable device with an adjustable magnification lens configured to align with a camera of an electronic device.

4. The system of claims 1 to 3, wherein the portable device further comprises: a lens slider configured to adjust a position of the magnification lens; and a device holder configured to secure the electronic device in alignment with the magnification lens.

5. The system of any one of claims 1 to 4, wherein the trained machine learning model is configured to classify the retinal fundus image as indicating presence of one or more specific types of retinal detachment.

6. The system of claim 5, wherein the one or more specific types of retinal detachment include rhegmatogenous retinal detachment, exudative retinal detachment, and tractional retinal detachment.

7. The system of any one of claims 1 to 6, wherein the system is further configured to: receive an ophthalmologist analysis of the retinal fundus image; and update the trained machine learning model based on the ophthalmologist analysis.

8. The system of any one of claims 1 to 7, wherein the system is further configured to flag cases that require immediate intervention for urgent ophthalmologistreview.

9. The system of claim 8, wherein the cases that require immediate intervention include macula-off retinal detachments.

10. The system of claim 8, wherein the system is further configured to identify and classify macula-on retinal detachments.

11. The system of any one of claims 1 to 10, wherein preprocessing the retinal fundus image comprises removing extraneous background information.

12. The system of any one of claims 1 to 11 , wherein the determination includes a probability score indicating likelihood of retinal detachment.

13. The system of any one of claims 1 to 12, wherein the trained machine learning model comprises a convolutional neural network.

14. The system of any one of claims 1 to 13, wherein the trained machine learning model is selected from the group consisting of ResNet, MobileNet, EfficientNet, DenseNet, and Inception-ResNet.

15. The system of any one of claims 1 to 14. wherein the trained machine learning model is configured to generate a heatmap highlighting regions of the retinal fundus image indicative of retinal detachment.

16. The system of any one of claims 1 to 15, wherein the trained machine learning model is configured to classify the retinal fundus image into multiple categories, including normal, early-stage retinal detachment, and advanced retinal detachment.

17. A method for detecting an ophthalmic condition, comprising: capturing a retinal fundus image of an individual using an imaging device; preprocessing the retinal fundus image to enhance retinal features; inputting the preprocessed retinal fundus image into a trained machine learning model, wherein the machine learning model is trained on a set of labeled retinal fundus images; generating, using the trained machine learning model, a determination of whether the retinal fundus image indicates presence of retinal detachment; and outputting the determination to a user interface.

18. The method of claim 17, wherein the ophthalmic condition is selected from the group consisting of retinal detachment, Central Retinal Artery Occlusion (CRAO), Endophthalmitis, and Papilledema.

19. The method of claims 17 or 18, wherein the trained machine learning model comprises a convolutional neural network.

20. The method of claims 17 to 19. further comprising: receiving an ophthalmologist analysis of the retinal fundus image; and updating the trained machine learning model based on the ophthalmologist analysis.

21. The method of any one of claims 17 to 20, wherein the trained machine learning model is configured to generate a heatmap highlighting regions of the retinal fundus image indicative of retinal detachment.

22. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform a method for detecting an ophthalmic condition, the method comprising: receiving a retinal fundus image captured by an imaging device; preprocessing the retinal fundus image to enhance retinal features; inputting the preprocessed retinal fundus image into a trained machine learning model; generating, using the trained machine learning model, a determination of whether the retinal fundus image indicates presence of retinal detachment; and outputting the determination to a user interface.

23. The non-transitory computer-readable storage medium of claim 22, wherein the ophthalmic condition is selected from the group consisting of retinal detachment, Central Retinal Artery Occlusion (CRAO), Endophthalmitis, and Papilledema.

24. The non-transitory computer-readable storage medium of claims 22 or 23, wherein the trained machine learning model is selected from the group consisting of convolutional neural networks, ResNet, MobileNet, EfficientNet, DenseNet, and Inception- ResNet.

25. A computer implemented method for detecting an ophthalmic disease, disorder or condition, the method comprising: generating a machine learning classifier that classifies medical data, including image data, into one of a plurality of classifications, wherein the machine learning classifier is generated by training a machine learning model, using ophthalmic images; obtaining an ophthalmic image of an individual; evaluating the ophthalmic image using the machine learning classifier to generate a determination of the ophthalmic disease, disorder, or condition, the determination having a sensitivity of at least 90% and a specificity of at least 90% when tested against an independent data set of at least 100 sample images; and providing the determination to the individual or a third party.

26. The method of claim 22, wherein the machine learning model can be comprised of, but is not limited to. Convolutional Neural Network, ResNet, MobileNet, LeNet, EfficientNet, AmoebaNet, Inception-ResNet, DenseNet, XCeption, any of InceptionNet, NasNet, SqueezeNet, ShuffleNet, GPipe, GoogleNet, SENet, or WideResNet.

27. The method of claim 22, wherein the ophthalmic image of the individual comprises a retinal fundus image.

28. The method of claim 24, wherein the retinal fundus image can be a fundus and / or ultra widefield fundus (UWF) image.

29. The method of claim 24, wherein obtaining the retinal fundus image comprises capturing the retinal image of the individual with an electronic device comprising a camera and a portable device comprising an imaging component wherein the portable device and the electronic device can be adjusted for positioning to align the imaging component with the camera.

30. The method of claim 22, wherein evaluating the ophthalmic image comprises uploading the ophthalmic image to the cloud network for remote analysis of the ophthalmic image using the machine learning model.

31. The method of claim 22, wherein the ophthalmic disease is a retinal disease or condition.

32. The method of claim 28, wherein the retinal disease or condition is retinal detachment, macula-on versus macula-off retinal detachment, rhegmatogenous retinal detachment, or non-rhegmatogenous retinal detachment such as exudative retinal detachment or fractional retinal detachment.

33. The method of claim 22, further comprising providing recommendations on the areas of the ophthalmic image where the retinal condition might be present and / or positioning the individual to slow dow n the progression of the disease.

34. The method of claim 22, w herein the determination of the retinal condition could be provided to the individual and / or third party on a w eb interface, mobile interface, the said electronic device or communicated via medical personnel.

35. A computer implemented system comprising: a server comprising at least one processor configured to generate a machine learning classifier that classifies medical data, including image data, into one of a pl urality of classifications, wherein the machine learning classifier is generated by training a machine learning model using ophthalmic images; an electronic device comprising at least one processor, a memory, a camera and anoperating system; a portable device comprising an imaging component, said portable device configured to receive and position the electronic device to align the camera with the imaging component; and a computer program stored in the memory of the electronic device, the computer program including instructions configured to, when executed by at least one processor of the electronic device, control the camera to capture an ophthalmic image of an individual, evaluate the ophthalmic image using the machine learning classifier to generate a determination of the ophthalmic disease, disorder, or condition, the determination having a sensitivity of at least 90% and a specificity of at least 90% when tested against an independent data set of at least 100 samples; and providing the determination to the individual or a third party.

36. The system of claim 32, wherein the machine learning model can be comprised of, but is not limited to. Convolutional Neural Network, ResNet, MobileNet, LeNet. EfficientNet. AmoebaNet, Inception-ResNet, DenseNet. XCeption. any of InceptionNet, NasNet, SqueezeNet, ShuffleNet, GPipe, GoogleNet, SENet, or WideResNet.

37. The system of claim 32, wherein the ophthalmic image of the individual comprises a retinal fundus image.

38. The system of claim 34, wherein the retinal fundus image can be a fundus and / or ultra widefield fundus (UWF) image.

39. The system of claim 34, wherein obtaining the retinal fundus image comprises capturing the retinal image of the individual with an electronic device comprising a camera and a portable device comprising an imaging component wherein the portable device and the electronic device can be adjusted for positioning to align the imaging component with the camera.

40. The system of claim 32, wherein evaluating the ophthalmic image comprises uploading the ophthalmic image to the server for remote analysis of the ophthalmic image using the machine learning model.

41. The system of claim 32, wherein the ophthalmic disease is a retinal disease or condition.

42. The system of claim 38, wherein the retinal disease or condition is retinal detachment, macula-on versus macula-off retinal detachment, rhegmatogenous retinal detachment, or non-rhegmatogenous retinal detachment such as exudative retinal detachmentor fractional retinal detachment.

43. The system of claim 32, wherein the computer program is further configured to provide recommendations on the areas of the ophthalmic image where the retinal condition might be present and / or positioning the individual to slow down the progression of the disease.

44. The system of claim 32, wherein the determination of the retinal condition could be provided to the individual and / or third party on a web interface, mobile interface, the said electronic device or communicated via medical personnel.

45. A computer implemented system comprising: a server comprising at least one processor configured to generate a machine learning classifier that classifies medical data, including image data, into one of a plurality of classifications, wherein the machine learning classifier is generated by training a machine learning model using ophthalmic images; a medical imaging device configured to capture an ophthalmic image of an individual; an electronic device operatively coupled to the medical imaging device, said electronic device comprising at least one processor, a memory, and an operating system; and a computer program stored in the memory of the electronic device, the computer program including instructions configured to, when executed by at least one processor of the electronic device, control the medical imaging device to capture an ophthalmic image of an individual, evaluate the ophthalmic image using the machine learning classifier to generate a determination of the ophthalmic disease, disorder, or condition, the determination having a sensitivity of at least 90% and a specificity' of at least 90% when tested against an independent data set of at least 100 samples; and providing the determination to the individual or a third party.

46. The system of claim 42, wherein the machine learning model can be comprised of, but is not limited to, Convolutional Neural Network, ResNet, MobileNet, LeNet, EfficientNet, AmoebaNet, Inception-ResNet, DenseNet, XCeption, any of InceptionNet, NasNet. SqueezeNet, ShuffleNet. GPipe. GoogleNet. SENet. or WideResNet.

47. The system of claim 42, wherein the ophthalmic image of the individual comprises a retinal fundus image.

48. The system of claim 44, wherein the retinal fundus image can be a fundus and / or ultra wi defield fundus (UWF) image.

49. The system of claim 44, wherein the medical imaging device is an ultrawidefield fundus camera or ultra widefield retinal camera.

50. The system of claim 42, wherein evaluating the ophthalmic image comprises uploading the ophthalmic image to the server for remote analysis of the ophthalmic image using the machine learning model.

51. The system of claim 42, wherein the ophthalmic disease is a retinal disease or condition.

52. The system of claim 48, wherein the retinal disease or condition is retinal detachment, macula-on versus macula-off retinal detachment, rhegmatogenous retinal detachment, or non-rhegmatogenous retinal detachment such as exudative retinal detachment or fractional retinal detachment.

53. The system of claim 42, wherein the computer program is further configured to provide recommendations on the areas of the ophthalmic image where the retinal condition might be present and / or positioning the individual to slow down the progression of the disease.

54. The system of claim 42, wherein the determination of the retinal condition could be provided to the individual and / or third party on a web interface, mobile interface, the said electronic device or communicated via medical personnel.

Images