Shaping the future of myopia: artificial intelligence for vitreoretinal complications of high and pathologic myopia.

Shaping the future of myopia: artificial intelligence for vitreoretinal complications of high and pathologic myopia - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Graefes Arch Clin Exp Ophthalmol . 2026 Feb 4;264(5):1255–1272. doi: 10.1007/s00417-025-07098-9 Search in PMC Search in PubMed View in NLM Catalog Add to search Shaping the future of myopia: artificial intelligence for vitreoretinal complications of high and pathologic myopia Yeabsira Mesfin Yeabsira Mesfin 1 School of Medicine, University of California San Francisco, San Francisco, CA US Find articles by Yeabsira Mesfin 1 , Anish Salvi Anish Salvi 2 School of Medicine, Stanford University, Palo Alto, CA US Find articles by Anish Salvi 2 , Leo Arnal Leo Arnal 2 School of Medicine, Stanford University, Palo Alto, CA US Find articles by Leo Arnal 2 , Curtis Langlotz Curtis Langlotz 2 School of Medicine, Stanford University, Palo Alto, CA US 3 Department of Radiology, Stanford University, Palo Alto, CA US Find articles by Curtis Langlotz 2, 3 , Vinit Mahajan Vinit Mahajan 2 School of Medicine, Stanford University, Palo Alto, CA US 4 Department of Ophthalmology, Byers Eye Institute, Stanford University, Palo Alto, CA US Find articles by Vinit Mahajan 2, 4 , Chase A Ludwig Chase A Ludwig 2 School of Medicine, Stanford University, Palo Alto, CA US 4 Department of Ophthalmology, Byers Eye Institute, Stanford University, Palo Alto, CA US 5 Department of Ophthalmology, Byers Eye Institute, 2370 Watson Court Office 100D, Palo Alto, CA 94303 US Find articles by Chase A Ludwig 2, 4, 5, ✉ Author information Article notes Copyright and License information 1 School of Medicine, University of California San Francisco, San Francisco, CA US 2 School of Medicine, Stanford University, Palo Alto, CA US 3 Department of Radiology, Stanford University, Palo Alto, CA US 4 Department of Ophthalmology, Byers Eye Institute, Stanford University, Palo Alto, CA US 5 Department of Ophthalmology, Byers Eye Institute, 2370 Watson Court Office 100D, Palo Alto, CA 94303 US ✉ Corresponding author. Received 2025 Aug 16; Revised 2025 Dec 6; Accepted 2025 Dec 21; Issue date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13091850  PMID: 41636834 Abstract Purpose The global impact of myopia extends far beyond individual ocular health, posing significant challenges to healthcare systems worldwide. Artificial intelligence (AI), particularly deep learning (DL) applied to ophthalmic imaging, offers a promising strategy to ease constraints posed by the myopia epidemic by detecting subtle structural changes early. Here we describe the current literature on AI for detecting retinal sequelae of myopia, including retinal detachments (RD), myopic macular degeneration (MMD), and myopic traction maculopathy (MTM), with attention to imaging modality and model task (classification vs. segmentation). Methods A literature search was conducted to identify studies using DL to detect RD, MMD, and MTM across ophthalmic imaging modalities (including OCT and fundus photography, and where available fluorescein angiography and ultrasonography). Results/findings We reviewed 28 studies that piloted DL models usingclassification and/or segmentation approaches for RD (10 studies), MMD (12 studies), and MTM (6 studies). Reported performance for RD ranged from area under the curve (AUC) 86-100%, accuracy 79.3-98.9%, sensitivity 77.1-97.6%, and specificity 79.7-100%. For MMD, performance ranged from AUC 86-100%, accuracy 85.3-99.8%, sensitivity 37.1-97.8%, and specificity 91.5-99.9%. For MTM, performance ranged from AUC 93.8-99.7%, accuracy 94.3-99.3%, sensitivity 74.5-98.4%, and specificity 84.8-99.7%. Across studies, there was substantial heterogeneity in case definitions, datasets, and evaluation methods, and external validation was inconsistently reported. Many earlier studies used CNN-based architectures, while more recent work increasingly incorporates transformer-based backbones and pretrained or foundation models. Conclusion Researchers have demonstrated excellent results for developing DL models that accurately classify and segment retinal pathologies associated with myopia. However, despite strong performance, additional work is needed to translate these models into clinical use, including robust external validation, calibration for clinical decision-making, and prospective evaluation, particularly for longitudinal prognostication of incident complications in pathologic myopia. Supplementary Information The online version contains supplementary material available at 10.1007/s00417-025-07098-9. Keywords: Deep learning, Artificial intelligence, Myopia, Pathologic myopia, Retinal detachment, Myopic macular degeneration, Myopic traction maculopathy Key Messages What is known Artificial intelligence (AI), particularly deep learning applied to ophthalmic imaging, offers a promising strategy to ease constraints posed by the myopia epidemic by detecting subtle structural changes early. What is new In this literature review, the studies we highlight reported strong performance for deep learning models that classify and segment retinal pathologies associated with myopia across common imaging modalities. However, additional work is needed for clinical implementation, including external validation and calibration, and to extend current models from cross-sectional detection to reliable longitudinal prognostication of disease course over time. Supplementary Information The online version contains supplementary material available at 10.1007/s00417-025-07098-9. Introduction on the myopia epidemic By 2050, nearly 50% of the world’s population is expected to be diagnosed with myopia [ 1 ]. While the etiology remains unclear, the causes of this growing epidemic are believed to be multifactorial, with studies finding an association with one’s family history, frequency of outdoor activities, prolonged near-work activities and extent of screen time [ 2 – 5 ]. The link between myopia and other ophthalmic conditions further complicates this epidemic. For instance, myopia has been shown to lead to a higher likelihood of cataracts, open-angle glaucoma, and retinal detachment (RD) [ 6 – 8 ]. Anatomically, the pathologic complications associated with myopia can be attributed to the axial elongation that accompanies this disorder, compromising the structural integrity of the retina and optic nerve head [ 9 ]. Accurate diagnosis and careful monitoring of patients with high myopia are therefore critical for improving long-term prognosis. Given these escalating risks and the anticipated surge in high myopia prevalence, it is evident that the global impact of myopia extends far beyond individual ocular health, posing significant challenges to healthcare systems worldwide. The expected increase in myopia prevalence will substantially strain ophthalmic care delivery, exacerbate existing bottlenecks, drive up healthcare costs, and result in higher rates of vision loss and blindness. With nearly one billion individuals projected to have high myopia by 2050, the incidence of sight-threatening vitreoretinal complications alone such as RD, myopic macular degeneration (MMD), and myopic traction maculopathy (MTM) will also rise [ 1 ]. Economically, the burden associated with managing myopia and its complications is projected to reach annual direct healthcare costs of $328 billion globally by 2030, rising further to $487 billion per year by 2050 [ 10 ]. This significant anticipated expenditure underscores the urgent need for innovative diagnostic and therapeutic approaches to alleviate both the economic and clinical burdens posed by this growing epidemic. Artificial intelligence (AI) offers a promising strategy to ease constraints posed by the myopia epidemic. By rapidly and accurately interpreting ophthalmic imaging, AI-driven diagnostic models can relieve bottlenecks in vitreoretinal care, allowing ophthalmologists to manage larger patient volumes more effectively. AI’s scalability and automation may also mitigate rising healthcare costs by streamlining workflows and reducing reliance on specialist resources. Crucially, AI’s capability to detect subtle disease early may significantly reduce vision loss linked to delayed diagnosis. Researchers have demonstrated excellent results developing deep learning (DL) models that accurately segment retinal pathology and classify complex vitreoretinal diseases. The next steps involve implementing these validated models clinically and developing predictive algorithms for proactive disease management. Herein, we review the current state of AI research for the primary vitreoretinal complications of myopia, MMD, MTM, and RRD, highlighting successful classification and segmentation algorithms, and discussing existing unmet needs and future research directions. Deep learning models for image analysis Highly accurate and therefore clinically feasible models for vitreoretinal complications of myopia have become possible due to advancements in DL methods. DL models are a subset of AI that utilize multilayered neural networks to simulate complex decision-making processes. Convolutional neural networks (CNNs) have historically been the dominant deep learning approach for medical image interpretation because they are well-suited to visual pattern recognition [ 11 ]. More recently, transformer-based vision models (e.g. Vision Transformers and hierarchical variants) and large pretrained 'foundation' backbones have become increasingly common, often outperforming CNNs when sufficient data or pretraining is available. However, many clinically deployed models and much of the legacy literature remain CNN-based or hybrid CNN-transformer systems. Structurally, the first level of CNNs consists of an input layer, composed of multiple groupings of nodes known as local receptor fields (LRF, Fig. 1 ). Each LRF processes the image and uses filters to detect specific features, generating feature maps that become the input for additional connected nodes within hidden layers. This hierarchical process is repeated in successive hidden layers, each layer pooling inputs from the preceding layer while applying additional filters. Finally, an output is generated by aggregating these feature vectors and interpreting them in relation to each other through a fully connected layer, functioning similarly to a regression model [ 11 ]. Fig. 1. Open in a new tab Schematic diagram of how convolutional neural networks analyze images to generate an output DL models can be designed to perform different types of image analysis tasks. Broadly, these tasks include classification, where the entire image is interpreted to recognize patterns associated with specific disease stages, and segmentation, which involves labeling images by pixels or voxels to precisely localize specific anatomical features such as lesions [ 12 , 13 ]. Other relevant tasks include object detection, in which models simultaneously localize and classify multiple structures or abnormalities within an image (e.g., bounding box), and regression, which aims to predict continuous outcomes (e.g., lesion size, visual acuity). In current practice, many systems are multitask (e.g. joint detection + segmentation) and increasingly multimodal, combining imaging with clinical variables or text reports using transformer-based architectures. Clinically, ophthalmologists frequently rely on imaging techniques such as optical coherence tomography (OCT), fundus photography, fluorescein angiography, and B-scan ultrasonography, integrating these with physical exam findings, laboratory results, and patient histories. Therefore, DL models employing classification, segmentation, and related tasks hold significant promise for the detection and analysis of retinal lesions in patients with pathologic myopia. Classification models Classification models analyze medical images globally and have demonstrated efficacy in diagnosing various pathologies. For instance, the ResNet-34 architecture, introduced by He et al. in 2015, is widely recognized for its effectiveness in medical image interpretation and classification tasks [ 14 ]. ResNet variants remain common baselines in ophthalmology, but newer backbones including Vision Transformers and modern convolutional/attention hybrids are increasingly used for state-of-the-art performance and for transfer learning from large pretrained models. ResNet architectures have shown promising results in the detection of skin cancer lesions, recognition of pathological tongue features, and the diagnosis of breast cancer from mammography images [ 15 – 17 ]. Classification models typically have lower computational complexity compared to segmentation models because they do not require detailed regional analysis of images, making them more efficient for rapid processing. Furthermore, the holistic approach of classification models allows for broader applicability across various types of medical images and diverse pathologies, potentially enhancing generalizability [ 18 ]. However, their global approach to image interpretation may reduce clinical precision, particularly when models are insufficiently trained or when pathology manifests in subtle, localized features [ 19 ]. They also often require substantial labelled training data, particularly when trained from scratch rather than initialized from large-scale pretraining. A notable limitation of classification models is the “black box” issue inherent to many artificial intelligence systems, referring to the opacity of their decision-making processes. This lack of transparency can make clinicians hesitant to rely on AI outputs for critical medical decisions, as the rationale behind the model’s conclusions is often unclear [ 20 , 21 ]. To address this challenge, visualization techniques such as Gradient-weighted Class Activation Mapping (Grad-CAM) have been developed. Grad-CAM provides heat maps highlighting regions within images that contributed significantly to the model’s predictions, offering some interpretability regarding how decisions are made. However, Grad-CAM has limitations. It does not fully elucidate the exact features or clinical reasoning behind the model’s predictions and may sometimes highlight irrelevant or ambiguous regions, leaving the clinician with partial explanations that require cautious interpretation [ 22 ]. Accordingly, recent work increasingly combines saliency-style visualization with model-agnostic attribution, counterfactual explanations, and clinically grounded evaluation (e.g., localization metrics and reader studies) to better align 'explanations' with what clinicians might consider meaningful. Given these limitations in clinical precision and interpretability inherent in classification models, segmentation models have emerged as a complementary approach, offering more detailed lesion analysis and improved localization of vitreoretinal complications associated with high myopia. Segmentation models Segmentation models provide a more granular and targeted analysis of images compared to classification models, enabling precise localization of specific retinal pathology. These models partition images into distinct regions or segments before identifying specific anatomical or pathological features relevant to diagnosis or analysis [ 23 ]. Historically, many medical segmentation systems were built on CNN encoder-decoder designs (e.g. U-Net-style architectures), while newer approaches increasingly incorporate transformer encoders or foundation segmentation models that can adapt with fewer task-specific labels. For instance, segmentation models have demonstrated efficacy in precisely identifying liver tumors and vasculature from computed tomography (CT) scans [ 24 ]. Similarly, segmentation-based approaches have improved the accuracy of detecting and delineating brain tumors in magnetic resonance imaging (MRI) scans [ 25 ]. Recently, vision-language foundation models (VLMs), such as Merlin (a 3D transformer-based model trained using CT scans, radiology reports, and electronic health records [EHRs]) have expanded segmentation capabilities to include zero-shot 3D semantic segmentation of complex anatomical structures, illustrating the significant potential of transformer architectures in medical imaging segmentation tasks [ 26 ]. In ophthalmology, the detailed, region-specific analysis offered by segmentation can enhance the accuracy of retinal lesion detection compared to classification models, which analyze images as a whole. Moreover, explicitly localizing and delineating image features, segmentation models provide clearer insights into the decision-making process of AI systems. This enhanced interpretability can help reduce clinicians’ hesitancy in adopting AI for clinical decisions. However, despite these advantages, segmentation models often have a higher computational cost and require extensive training with large, annotated datasets to reliably learn to segment clinically relevant features from medical images [ 23 ]. Annotating large image datasets is resource-intensive and time-consuming, posing a practical challenge. Furthermore, because segmentation models are typically highly specialized, their performance may not generalize well across diverse datasets or imaging modalities [ 27 ]. Consequently, developing and maintaining fully bespoke algorithms for each modality or institution can be commercially and practically challenging[ 27 ]. This has motivated growing interest in foundation models, domain adaptation, and lightweight fine-tuning strategies that aim to improve portability across settings. Ultimately, the limited availability of sufficiently diverse, annotated training datasets remains a major obstacle for widespread adoption of segmentation models. Nevertheless, segmentation-based deep learning methods have already demonstrated considerable promise in diagnosing retinal complications associated with pathologic myopia. AI models in the detection of myopia sequelae The earliest clinically actionable signs of pathologic myopia, including retinal detachments, myopic macular degeneration, and myopic traction maculopathy, can often be detected with careful retinal examination and multimodal imaging. Clinically, OCT and fundus imaging are central to evaluation, and DL models have shown strong performance for automated detection and feature extraction from these modalities. With the growing prevalence of myopia, these tools can support scalable screening and longitudinal monitoring by detecting subtle structural changes over time, enabling earlier referral and more consistent follow-up in high risk eyes. Retinal detachments Retinal detachments (RD) are characterized as a separation of the neurosensory retina from the underlying retinal pigment epithelium layer, typically due to a retinal break with accumulation of subretinal fluid [ 28 ]. Axial elongation in the setting of high myopia is among the largest contributing risk factors towards detachments by mechanically stretching the retina and vitreous [ 29 ]. This anatomical distortion compromises the retina’s structural integrity and predisposes it to small breaks that allow the seepage of fluid between the retinal layers, at times culminating in a full detachment [ 30 ]. Not only does myopia increase one’s risk of RD, but it does so in a dose-dependent manner such that patients with higher refractive errors have higher likelihoods of detachment [ 31 ]. This represents a growing concern for patients with high myopia, warranting advancements in the diagnostic and therapeutic measures to mitigate this risk. Classification models in the detection of RD RD is recognizable on imaging by anatomical features such as retinal elevation with subretinal fluid, and, when visible, associated retinal breaks; additionally, lattice degeneration is an important risk marker ( Supplemental Fig. 1 ). Consequently, DL classification algorithms have been developed to diagnose RD by detecting these characteristic lesions (Table 1 ) [ 32 – 41 ]. In two such studies by Zhang et al. and Ohsugi et al., both utilized ultra-widefield fundus imaging, capturing high-resolution images without pupil dilation, making it accessible for use by non-specialist healthcare providers [ 32 , 33 ]. Table 1. Literature review of deep learning models in the detection of retinal detachment and its lesion subtypes Author, Year Architecture Dataset Size Lesion Subtypes AUC Accuracy Sensitivity Specificity F1 Score Precision IOU DSC Ref Zhang et al., 2021 ResNext-50 189 fundus images Lattice Degeneration 0.880 79.3% 77.10% 79.70% -- -- -- -- Retinal Breaks 0.953 92.0% 87.50% 92.40% -- -- -- -- [ 38 ] Retinal Detachments 1 98.9% 87.50% 100% -- -- -- -- Ohsugi et al., 2017 Unspecified CNN 831 fundus images Retinal Detachments 0.988 -- 97.60% 96.50% -- -- -- -- [ 39 ] Yadav et al., 2024 MultiResU-Net 317 fundus images Retinal Detachments -- -- -- -- 93.3 -- 88.4 90.8 [ 40 ] Caki et al., 2025 ResNet-50 + Unet 213 US images Retina/Choroid -- -- 85.20% -- 84.7 85.70% -- -- [ 41 ] Sclera -- -- 81.30% -- 78.3 76.40% -- -- Optic Nerve Sheath -- -- 88.40% -- 88.2 89.50% -- -- Retinal Detachments -- 98.1% 96.00% -- 96.5 97.60% -- -- Christ et al., 2024 EfficientNet 2849 fundus images Retinal Breaks 0.913 84.4% 86.30% 82.80% -- 80.80% -- -- [ 42 ] Retinal Detachments 0.972 92.3% 86.90% 95.40% -- 91.60% -- -- Retinal Lesions 0.975 94.3% 91.90% 97.90% -- 98.50% -- -- Zhou et al., 2022 ResNet-50 554 fundus images Recurrent Retinal Detachment 0.935 91.2% 84.00% 92.90% -- -- -- -- [ 43 ] Inception-ResNet-V2 Recurrent Retinal Detachment 0.944 90.6% 89.00% 87.80% -- -- -- -- Li et al., 2020 InceptionResNet-V2 11,087 fundus images Retinal Detachments 0.989 98.9% 96.10% 99.60% -- -- -- -- [ 44 ] Macula-on Detachments 0.975 91.7% 93.80% 90.90% -- -- -- -- Rashid et al., 2025 ResNet-50 70,000 OCT images Retinal Detachments 0.987 97.8% 97.40% 98.60% 97.6 -- -- -- [ 45 ] Chen et al., 2021 ResNet-50 + Unet 4521 US images Normal vs. Abnormal Retina 0.950 94.0% 94% 95% -- -- -- -- [ 46 ] Vitreous Detachments 0.890 90.0% 88% 91% -- -- -- -- Vitreous Hemorrhage 0.860 92.0% 79% 94% -- -- -- -- Retinal Detachments 0.930 94.0% 92% 95% -- -- -- -- Other Retinal Lesions 0.910 91.0% 92% 91% -- -- -- -- Younis et al., 2025 YOLOv8 419 OCT images Choroidal Neovascularization, Macular Edema, Drusen, Normal -- -- 97.00% -- -- 96.80% -- -- [ 47 ] Open in a new tab Zhang et al. for example, developed binary classification models for the automated detection of RD, retinal tears (RT), and lattice degeneration, enhancing the ResNet50 architecture with a squeeze-and-excitation module. This addition allowed their model to recognize correlations between feature channels in fundus images and selectively amplify lesion-specific features. Additionally, Zhang et al. implemented two processing strategies when testing their algorithm’s performance: an original resizing method, in which each ultra-widefield fundus image was rescaled to fit the model’s input, and a cropping method, in which the images were partitioned into equally sized segments that were individually fed into the model. Ultimately, their AI model performed well in the detection of lattice, RT, and RD, achieving an area under the curve (AUC) of 0.888, 0.843, and 1.000 with the original resizing method and 0.841, 0.953, and 0.979 with the cropping method, respectively [ 32 ]. Through the original resizing method, a sensitivity of 77.1%, 62.5%, and 87.5%, and a specificity of 79.7%, 95.3%, and 100.0% was observed for the identification of each lesion, respectively. Meanwhile, through the cropping method, a sensitivity of 82.9%, 87.5%, and 81.2%, and a specificity of 75.2%, 92.4%, and 95.3% was achieved, respectively. Overall, their results suggest that with their model, the original resizing method is ideal for detecting lattice and RD, while the cropping method is optimal for detecting RT. However, these performance metrics are confounded by the small size of their test dataset when classifying each lesion, with only 57 images utilized for lattice, 16 for RT, and 16 for RD. Furthermore, their model was tested on fundus images of only tessellated eyes, which represent a narrow subset of myopic patients, limiting the generalizability of their results. By contrast, Ohsugi et al. designed a custom CNN specifically for binary classification of only RD with a significantly larger test dataset and broader patient population. Their study’s processing strategy mirrored the original resizing method utilized by Zhang et al. but did not incorporate a cropping method. Ohsugi et al.’s architecture, also trained on ultra-widefield fundus images, showed excellent performance when detecting RD as well, displaying a sensitivity of 97.6%, specificity of 96.5%, and AUC of 0.988 [ 33 ]. These indices outperform Zhang et al.’s cropping method for RD detection, and, although are slightly inferior to Zhang et al.’s original resizing method, benefit from having a substantially larger test set of 166 images, as compared to only 16 images. Ohsugi et al.’s high diagnostic accuracy on ultra-widefield fundus images facilitates remote interpretation by DL algorithms, which could significantly reduce the reliance on ophthalmologists, particularly benefiting rural and underserved communities. Collectively, both Zhang et al. and Ohsugi et al. illustrate that integrating DL classification models with ultra-widefield imaging devices such as Optos can significantly improve diagnostic accuracy and accessibility for retinal detachment, particularly in regions with limited specialist availability. Segmentation models in the detection of RD Segmentation models have also performed well in the detection of RD by segmenting lesions indicative of detachments. Yadav et al. designed the DCA-MultiResU-Net algorithm, a DL model that expands upon the features of the MultiResU-Net architecture (Table 1 ) [ 34 ]. Their model integrates dilated convolutions and channel attention modules to segment lesions related to RD in fundus images. By using dilated convolution, their model can capture larger spatial features of fundus images without increasing the number of parameters. Meanwhile, the application of a channel attention module enhances their model’s capacity to focus on the most relevant channels of detachment in fundus images, reducing the weight of irrelevant channels. In doing so, their system’s segmentation managed to outperform other U-Net and ResU-Net algorithms in properly delineating RD lesions, exhibiting a dice similarity coefficient (DSC) of 90.8%, intersection over union (IOU) of 88.4%, and F1 score of 93.3% [ 34 ]. Notably, Yadav et al. only highlight the results of their model’s ability to segment RD lesions but fail to report the performance indices when classifying RD against normal fundus images. This limits the practical implications of their DCA-MultiResU-Net system in a clinical setting since it is designed to localize lesions only after an RD diagnosis is confirmed. Meanwhile, Caki et al., also generated a DL model for the segmentation and diagnosis of RD but from B-Scan ocular ultrasonography (USG) images (Table 1 ) [ 35 ]. Their TransUNet framework is a hybrid model that integrates convolutional and transformer-based approaches. Rather than applying a traditional UNet architecture, their model builds off it with a ResNet-50 backbone for more granular feature extraction from USG images. This was further enhanced with a ViT module to better contextualize the extracted features in relation to the full image. In doing so, their model managed to outperform other standard networks such as ResNet-50 and MobileNetV3, exhibiting a F-score of 96.5%, precision of 97.6%, and recall of 96.0% when diagnosing RD from USG images [ 35 ]. The model’s 98.1% accuracy when classifying images of posterior vitreous detachments as non-RD further substantiates its performance compared to Yadav et al. Furthermore, regarding the segmentation of key anatomical structures, the TransUNet model managed to correctly segment the retina/choroid, sclera, and optic nerve sheath in both normal and diseased eyes with an F-score of 84.7%, 78.3%, and 88.2%, respectively [ 35 ]. Caki et al.’s algorithm surpasses that of Yadav et al.’s because it integrates both segmentation and classification modules to facilitate not only the precise localization of retinal structures but also the diagnosis of RD. This contrasts the singular use of segmentation without classification observed in Yadav et al. However, despite the broader capabilities of Caki et al.’s model, it segments normal anatomical structures rather than defects pertinent to the RD diagnosis. This may limit its translational impact since, in contrast to Yadav et al., it does not offer pixel-level delineations to localize detachments since their segmentation does not include such defects. Importantly, by expanding the application of such segmentation models towards USG represents yet another means by which DL models can improve the accessibility of ophthalmic care. Given the portability and low cost of ocular ultrasounds, this approach towards AI-driven medicine can facilitate more equitable access to care in rural communities that lack ophthalmologists. The unmet needs and implications of AI in the detection of RD Retinal detachment represents a sight-threatening complication of pathologic myopia that requires timely diagnosis and intervention [ 28 ]. While DL models have been efficacious in the diagnosis of RD, few are capable of the prognostication and risk stratification of RD. The lack of multi-modal and multi-center datasets for model training contributes to this limitation. Most deep learning models are trained on datasets from a limited number of medical institutions. Additionally, these datasets commonly incorporate images from a single imaging modality. This homogenous approach for training can limit the generalizability and integration of AI models for diverse populations. Simultaneously integrating clinical data from different sources, such as OCT, fundus photography, and ultrasonography, into the development of a single model can significantly enhance the performance of AI. Similarly, using multi-center validation can diversify how these systems are trained to improve their applicability towards a larger number of communities. Ultimately, both approaches could encourage a more holistic rationalization for a diagnosis since these deep learning models can rely on multiple sources to substantiate their outputs. Additionally, a more diverse training protocol has prognostic value. Increasing the number of image modalities that AI models analyze increases the number of biomarkers that these models interpret before deciding on a diagnosis. Thus, with a greater reliance on more inputs, the systems may be able to better predict a patient’s likelihood of detachment. For instance, retrospectively amplifying the weight of channels associated with certain biomarkers in the pre-detachment images of RD patients can help train AI models to correlate biomarkers with the likelihood of RD. This can facilitate the early prediction and detection of RD for more targeted interventions and encourage more informed decisions by ophthalmologists. Additionally, improving the prognostic and predictive abilities of DL models in this manner can also encourage more personalized care for patients. For example, these models can also help identify patients at greatest risk of detachment, encouraging earlier interventions that can prevent detachment. Similarly, they can also be trained to assess a patient’s likelihood of anatomical success following vitrectomies in the pre-operative period. This can help specialists gauge which patients would benefit most from certain interventions. Lastly, multi-center training sets can also simplify the integration of these systems into clinical practice. The complexity of these models and the ambiguity surrounding their internal function can create hesitancy towards adopting AI into clinics and relying on it for medical decision making. However, training these models with different imaging modalities and with images from different medical institutions can improve their approachability and accessibility. For instance, if AI models have the capacity to diagnose lesions from different imaging devices, disadvantaged communities with limited access to OCT scans can still utilize these models to diagnose lesions based off more readily available imaging devices. This can also create for a more user-friendly experience for clinics that are more dependent on devices that are not common practice. In essence, by familiarizing newer AI models to diverse training methods, newer deep learning models can be integrated into clinic more seamlessly. Myopic macular degeneration Myopic maculopathies such as myopic macular degeneration (MMD) are among the sight-threatening complications facing patients with high myopia [ 42 ]. Axial elongation in highly myopic patients results in choroidal thinning, reducing choroidal perfusion and oxygen delivery to the neurosensory retina, ultimately leading to progressive macular degeneration and vision loss [ 42 ]. On examination, MMD can be characterized by retinal findings such a tessellated fundus, peripapillary atrophy, posterior staphylomas, lacquer cracks, choroidal neovascularization, and Fuchs spots [ 43 ]. Thus, OCT scans and fundus photography are integral to supporting the diagnosis ( Supplemental Fig. 2 ). Furthermore, the variety of retinal lesions associated with MMD can allow deep learning models to detect the lesions efficiently. Current modalities for diagnosing pathologic myopia include OCT and fundus photography, both of which assist in classifying and staging the disorder [ 44 ]. For example, Ohno-Matsui et al. established the International Photographic Classification and Grading System for Pathologic Myopia (META-PM), categorizing patients based on retinal lesions from category 0 (“no myopic retinal lesions”) through category 4 (“macular atrophy”) and additionally highlighting “plus” lesions such as lacquer cracks, choroidal neovascularization, and Fuchs spots [ 45 ]. Wong et al. expanded upon the META-PM criteria by also grading optic disc abnormalities, including optic disc tilt, peripapillary atrophy, and peripapillary intrachoroidal cavitation [ 46 ]. More recently, Ruiz-Medrano et al. proposed the ATN classification system based on atrophic, tractional, and neovascular retinal changes [ 43 ]. Historically, variability in classification systems has led to the use of different diagnostic schema across deep learning studies. However, the International Myopia Institute (IMI) is now actively working toward establishing universal definitions and standardized classification criteria for pathologic myopia and myopic macular degeneration, facilitating consistent diagnosis across clinical practice and research [ 47 , 48 ]. Standardized definitions are particularly important for AI because inconsistent grading and label noise can substantially limit generalizability. The following sections highlight existing deep learning models designed to classify and segment MMD, showcasing their strengths and identifying areas requiring further refinement. Classification models in the detection of MMD DL algorithms have also been developed to diagnose MMD (Table 2 ) [ 49 – 60 ]. Wu et al., demonstrated the clinical applications of DL models in the diagnosis of MMD. Wu et al., enhanced the ResNet-34 architecture with a multi-branch model, a spatial pyramid pooling module, and an attention module to refine feature extraction from fundus imaging [ 49 ]. Then, through the application of a binary classification algorithm paralleling the ATN system, they trained their model to diagnosis myopic maculopathies. Their proposed Resnet-34 framework demonstrated an area under the curve (AUC) of 0.969, 0.895, and 0.936 and an accuracy of 92.38%, 85.34%, and 94.21% for the detection of atrophic, tractional, or neovascular changes, respectively, in the retina based off fundus photography. The performance of their deep learning model was found to even surpass that of two attending ophthalmologists and to be comparable to two retina specialists. Of note, by assigning ATN labels derived from OCT to fundus images, a strength of the model designed by Wu et al. also lies in its ability to infer the presence of atrophy, traction, and neovascularization from fundus images alone. However, its binary classification of ATN lacks granularity, which may limit the model’s potential for disease monitoring. Table 2. Literature review of deep learning models in the detection of myopic macular degeneration and its lesion subtypes Author, Year Architecture Dataset Size Lesion Subtypes AUC Accuracy Sensitivity Specificity F1 Score Precision IOU DSC Ref Wu et al., 2022 ResNet-34 CNN 1853 fundus images Retinal Atrophy 0.969 92.3% 92.2% 91.5% 91.6 91.1% -- -- [ 56 ] Macular Hole/Detachment 0.895 85.3% 72.5% 96.6% 75.8 83.6% -- -- Myopic Choroidal Neovascularization 0.936 94.2% 85.9% 98.1% 88.5 91.7% -- -- Zhao et al., 2024 ResNet-101 CNN 581 fundus images Normal fundus, Tessellated fundus, Early pathologic myopia, Advanced pathologic myopia, and Other diseases. 0.999 99.7% -- -- 99.7 -- -- -- [ 57 ] Tang et al., 2022 ResNet-50 1395 fundus images Pathologic Myopia 0.998 93.7% 96.7% 99.2% -- -- -- -- [ 58 ] No Myopic Lesion -- -- 92.9% 99.5% -- -- -- -- Tessellated Fundus -- -- 97.8% 94.6% -- -- -- -- Diffuse Chorioretinal Atrophy -- -- 89.8% 96.7% -- -- -- -- Extramacular Patchy Chorioretinal Atrophy -- -- 79.2% 97.7% -- -- -- -- Macular Patchy Chorioretinal Atrophy -- -- 50.0% 98.7% -- -- -- -- DeepLabv3+ Optic Disc -- -- 96.6% -- 94.6 92.7% 89.8 -- Peripapillary Atrophy -- -- 89.7% -- 90.0 90.3% 81.8 -- Lacquer Cracks -- -- 20.1% -- 23.8 29.1% 13.7 -- Diffuse Atrophy -- -- 87.4% -- 88.1 88.8% 78.7 -- Lacquer Cracks -- -- 85.3% -- 80.4 75.9% 67.2 -- Ali et al., 2024 MobileNetV3 396 fundus images Normal fundus, Tessellated fundus, Chorioretinal atrophy, Macular atrophy, Patchy chorioretinal atrophy, Fuch’s spot, lacquer cracks. 0.940 94.0% 97.5% -- 94.0 92.0% 96.0 -- [ 59 ] Du et al., 2021 EfficientNet 7020 fundus images Macular Atrophy 0.982 97.4% 85.1% 98.3% -- -- -- -- [ 60 ] Diffuse Atrophy 0.97 89.3% 84.4% 94.5% -- -- -- -- Patchy Atrophy 0.978 94.9% 87.2% 96.0% -- -- -- -- Choroidal Neovascularization 0.881 91.1% 37.1% 97.3% -- -- -- -- Lu et al., 2021 ResNet-18 CNN 3210 fundus images Normal fundus, Tessellated fundus, Chorioretinal atrophy, Macular atrophy, Patchy chorioretinal atrophy 0.979 96.7% -- -- -- -- -- -- [ 61 ] Choroidal neovascularization -- 97.0% 97.3% 97.0% -- -- -- -- Fuch’s spot -- 97.1% 97.8% 97.1% -- -- -- -- Lacquer Cracks -- 99.4% 68.4% 99.5% -- -- -- -- Park et al., 2022 EfficientNet-B4 NA Posterior staphylomas, Diffuse choroidal atrophy, Patchy chorioretinal atrophy, Macular atrophy, 0.980 95.0% 93.0% 96.0% -- -- -- -- [ 62 ] Choi et al., 2021 ResNet-50 1380 OCT images High Myopia Retinal Lesions 0.990 -- -- -- -- -- -- -- [ 63 ] High Myopia Retinal Lesions 0.970 -- -- -- -- -- -- -- High Myopia Retinal Lesions 0.860 -- -- -- -- -- -- -- Ye et al., 2021 ResNetst-101 CNN 2342 OCT images Macular Choroidal Thinning 0.927 -- 90.5% 88.7% -- -- -- -- [ 64 ] Bruch Membrane Defect 0.938 -- 88.9% 84.8% -- -- -- -- Subretinal Hyperreflective Material 0.927 -- 73.9% 91.3% -- -- -- -- Myopic Traction Maculopathy 0.974 -- 92.8% 90.5% -- -- -- -- Dome Shape Macula 0.955 -- 74.5% 94.0% -- -- -- -- Rauf et al., 2021 2–128 N-0D CNN 400 fundus images Pathologic Myopia 0.985 95.0% -- -- -- -- -- -- [ 65 ] Wang et al., 2022 PyTorch + EfficientNet-B8 10,347 fundus images Normal fundus or Mild tessellation 0.980 94.9% 93.1% 97.6% -- -- -- -- [ 66 ] Severe Tessellation 0.950 93.2% 92.9% 93.3% -- -- -- -- Early-Stage Pathologic Myopia 0.990 98.1% 90.8% 98.9% -- -- -- -- Advanced Stage Pathologic Myopia 1 99.8% 96.8% 99.9% -- -- -- -- Hemeling et al., 2021 ResNet-28 3350 fundus images Pathologic Myopia 0.987 -- -- -- -- -- -- -- [ 67 ] Unet++ Optic Nerve Head -- -- -- -- 98.7 -- -- 93.0 Atrophy -- -- -- -- 91.4 -- -- 80.0 Detachment -- -- -- -- 70.6 -- -- 80.7 Open in a new tab AUC : area under curve; IOU : intersection over union; DSC : dice similarity coefficient; OCT : optical coherence tomography Zhao et al. address this by integrating a direct classification model using the ResNet-101 architecture to extract features from fundus images; this architecture is deeper but structurally similar to ResNet-34, another deep learning framework commonly used for image classification. Zhao et al., applied two independent multi-classification schemas for the diagnosis of myopic maculopathies (Table 2 ) [ 50 ]. The first is an extension of the META-PM system but distinguishes pathologic myopia by categorizing images into A0 (normal or mildly tessellated fundus; termed 'normal or mild lattice basement membrane' by the authors), A1 (severely tessellated fundus; termed 'severe lattice basement membrane'), A2 (early pathologic myopia), A3 (late pathologic myopia), and Other. In contrast to Wu et al., their second classification schema defines eight additional categories, including four subcategories of early pathologic myopia relating to chorioretinal myopic degeneration and four subcategories of late pathologic myopia regarding the extent of patchy atrophy. In doing so, the deep learning framework designed by Zhao et al., demonstrated high performance in the diagnosis of pathologic myopia. When classifying pathologic myopia, the first schema had an AUC of 0.9896 and accuracy of 98.86, while the second exhibited an AUC of 0.9998 and accuracy of 99.67%. More so, these outcomes were also found to surpass those of other DL models as well, such as ViTs, EfficientNet, and VGG16. Notably, the broader classification schema in Zhao et al. allows their model to classify fundus images with additional granularity than the model generated by Wu et al. In doing so, their model outperforms that of Wu et al., in its potential to monitor the progression of early- to late-stage complications of pathological myopia from fundus images. However, OCT offers greater diagnostic accuracy when classifying the atrophic, tractional, and neovascular changes associated with pathologic myopia. Thus, while Zhao et al.’s study provides better granularity when detecting retinal lesions, Wu et al.’s integration of ATN labels derived from OCT images may confer greater prognostic value. Regardless, both studies underscore the clinical implications of classifications algorithms in diagnosing MMD. Segmentation models in the detection of MMD The incorporation of segmentation has also extended into the detection of myopic maculopathies. Tang et al., for instance, enhanced the capabilities of the ResNet system by integrating a segmentation architecture known as DeepLabv3+ (Table 2 ) [ 51 ]. Subsequently, their team manually annotated fundus photographs to highlight anatomical features pertinent to myopia, creating a training set for their model to segment the photographs properly. These manually annotated ocular features allowed the system to independently segment key anatomical and pathological findings such as the optic disc, peripapillary atrophy, diffuse chorioretinal atrophy, patchy chorioretinal atrophy, macular atrophy, lacquer cracks, and choroidal neovascularization, prior to classifying images according to the META-PM criteria. This co-dependent model, which integrates segmentation features, correctly classified 93.7% of fundus photographs, an improvement from 90.76% observed in the basic ResNet architecture. It also diagnosed pathologic myopia from non-pathologic myopia with an AUC, sensitivity, and specificity of 0.9980, 96.67%, and 99.15%, respectively. Of note, the sensitivity-specificity of their co-dependent model appeared to decrease with more severe stages of myopic maculopathy, suggesting that their model may be more optimized for early stages of this disorder. Moreover, their DeepLabv3 + segmentation architecture performed strongly in the pixel-level delineation of optic discs (F1: 94.6%; IOU: 89.8%), peripapillary atrophy (F1: 90.0%; IOU: 81.8%), diffuse atrophy (F1: 88.1%; IOU: 78.7%), and macular atrophy (F1: 80.4%; IOU: 67.2%) but poorly in the delineation of lacquer cracks (F1: 23.8%; IOU: 13.7%). Overall, Tang et al.’s integration of segmentation into their model encourages more reliable staging of MMD through nuanced classification. Meanwhile, Ali and Raut engineered a mobile app building off the Lightweight MobileNetV3 architecture with enhanced spatial attention and squeeze-excitation modules for the binary classification of pathologic myopia (Table 2 ) [ 52 ]. By incorporating these modules and training on annotated fundus image datasets, their model gained improved capability to identify and segment key pathological features, including chorioretinal atrophy, lacquer cracks, macular atrophy, and Fuchs spots. Ultimately, their app demonstrated high robustness when analyzing images from different heterogenous datasets, reporting an AUC of 0.94 in the binary classification of pathologic myopia. These results represented a notable improvement compared to a baseline model without spatial attention or squeeze-excitation modules, which achieved an AUC of 0.65 on corresponding tasks. Their segmentation architecture yielded an average IOU and average F1 score of 96.0% and 97.0%, respectively, and though such results are suggestive of optimal performance, Ali and Raut do not stratify their model’s segmentation performance by specific pathologic lesions. This limits the clinical implications of their model when compared to Tang et al., whose algorithm not only outperforms it in the overall classification of MMD but also reports more nuanced performance metrics for the segmentation of specific lesions. Despite this, the mobile architecture designed by Ali and Raut uniquely underscores not only the feasibility of these systems in clinical practice but also their portability and potential for increased accessibility of care. Prediction models in MMD prognosis Historically, DL models were primarily designed for classification or segmentation tasks in the diagnosis of MMD. However, recent advancements have extended their applications to predictive modeling, enabling clinicians to forecast disease progression and visual prognosis in patients with pathologic myopia. Chen et al. recently developed an interpretable machine learning algorithm designed to predict the risk of MMD progression over a ten-year period. Their model integrated clinical variables, including axial length, subfoveal choroidal thickness (SFCT), best-corrected visual acuity (BCVA), and anterior chamber depth, alongside demographic factors such as age and gender, achieving high predictive accuracy (AUC of 0.87) (Table 2 ) [ 61 ]. Similarly, epidemiologic models have also identified risk factors for progression of MMD, including older age, higher myopic refractive errors, and longer axial length [ 62 ]. Despite these promising outcomes, predictive models still require further validation to ensure their clinical applicability across diverse patient populations and varied imaging modalities. The generalizability of these models remains a significant concern, as differences in patient demographics, imaging protocols, and longitudinal data availability across clinical sites may influence predictive accuracy. Additionally, widespread integration of these predictive models into routine clinical practice has yet to be fully realized, highlighting the need for ongoing research to refine model interpretability, usability, and robustness in real-world clinical settings. Continued validation studies and further development of these prediction algorithms are essential to establish their role in personalized management and early intervention for patients at risk of developing severe MMD. The unmet needs and implications of AI in the detection of MMD While significant strides have been made in the classification, segmentation, and predictive modeling of MMD, several critical areas of clinical practice still present unmet needs. One major limitation is the early detection of subtle macular changes preceding irreversible MMD. In early stages, pathologic myopia can insidiously progress toward MMD, with axial elongation causing subtle retinal thinning and early atrophic changes that precede clearly identifiable lesions on traditional imaging methods [ 43 , 63 ]. The nuanced nature of these early pathological features limits accurate annotation, thus restricting training datasets and limiting the capability of DL models to reliably detect subtle signs of early-stage MMD. This challenge is further complicated by the clinical overlap between early-stage MMD and other ocular conditions, such as age-related macular degeneration (AMD), which can also present with choroidal neovascularization (CNV), making differentiation difficult for AI-based models [ 64 ]. Moreover, although accurate lesion identification using AI models is well-established, their applications toward therapeutics in MMD remain limited. For instance, CNV-induced abnormal angiogenesis in MMD necessitates treatments like anti-vascular endothelial growth factor (VEGF) injections [ 65 ]. AI-driven identification of high-risk lesions could theoretically guide more precise decisions on therapeutic interventions. Similarly, earlier detection and risk stratification could enable targeted surveillance and timely intervention for complications such as myopic CNV, rather than relying on uniform follow-up schedules [ 66 ]. Nevertheless, uncertainty surrounding AI-based decision-making processes, often described as the “black box” issue, continues to inhibit clinical confidence and integration into personalized care strategies [ 21 , 61 ]. Despite recent efforts toward improved interpretability in AI systems, hesitancy persists, reflecting a critical barrier to the widespread clinical adoption of these advanced models. Ultimately, addressing these remaining challenges requires continued refinement of AI algorithms, further validation across diverse clinical settings, and improved transparency and interpretability of model decision-making processes to enhance clinician trust and facilitate the integration of AI into personalized patient care. Myopic traction maculopathy Myopic traction maculopathy (MTM) is an umbrella-term that characterizes the tractional changes to the fovea and macula in highly myopic eyes, including foveoschisis, maculoschisis, foveal detachments, lamellar holes and full-thickness macular holes [ 67 ]. Its pathophysiology is multifactorial but largely related to how preretinal forces due to posterior vitreous detachments, vitreomacular traction, or epiretinal membranes and subretinal forces due to scleral deformity overwhelm the elasticity of the retina [ 68 ]. The degeneration of the choroid and retinal pigment epithelium is thought to also increase the retina’s susceptibility to these tractional forces [ 68 ]. Thus, the anatomical stretching of these retinal layers seen in patients with high myopia represents a large predisposition to MTM. OCT imaging remains the gold standard for diagnosing MTM because of how precisely it can visualize the morphological changes associated with this disorder. Consequently, DL models that can interpret these images for disease classification and detection have been proposed as well. Classification models in the detection of MTM The morphological changes associated with MTM subtypes are best distinguished on OCT imaging ( Supplemental Fig. 3 ). For instance, while the progression of foveoschisis localizes to an intact fovea other subtypes such as maculoschisis occur extra-foveally or are indicated by a distinct separation or break in the retinal layers, such as with foveal detachments or macular holes [ 67 , 69 ]. These diagnostically distinct deformities can facilitate the training of DL models, optimizing the utilization of direct classification systems for MTM grading (Table 3 ) [ 57 , 70 – 74 ]. Table 3. Literature review of deep learning models in the detection of myopic traction maculopathy and its lesion subtypes Author, Year Architecture Dataset Size Lesion Subtypes AUC Accuracy Sensitivity Specificity F1 Score Precision IOU DSC Ref. Chen et al., 2023 FIT-Net 9984 OCT images Retinoschisis 0.975 94.3% 92.9% 95.6% 94.2 -- -- -- [73] Macular Hole 0.997 98.7% 97.5% 99.0% 97.3 -- -- -- Retinal Detachment 0.996 98.8% 96.3% 99.7% 97.6 -- -- -- Huang et al., 2023 ResNet-34 CNN 2837 OCT images No Myopic Traction Maculopathy 0.993 97.5% 96.7% 98.4% -- -- -- -- Extrafoveal Maculoschisis 0.996 97.7% 92.3% 98.3% -- -- -- -- [ 77 ] Inner Lamellar Macular Hole 0.984 98.2% 91.7% 98.9% -- -- -- -- Outer Foveoschisis 0.996 99.2% 97.8% 99.4% -- -- -- -- Foveal Detachment 0.998 99.2% 96.7% 99.2% -- -- -- -- Ye et al., 2021 ResNeSt101 CNN 2342 images Myopic Traction Maculopathy 0.974 -- 92.8% 90.5% -- -- -- -- [ 64 ] Macular Choroidal Thinning 0.927 -- 90.5% 88.7% -- -- -- -- Bruch Membrane Defects 0.938 -- 88.9% 84.8% -- -- -- -- Dome Shaped Macula 0.955 -- 74.5% 94.0% -- -- -- -- Subretinal Hyperreflective Material 0.927 -- 73.9% 91.3% -- -- -- -- Jiang et al., 2024 HyFormer 2162 OCT images Internal Limiting Membrane Defects -- -- 79.4% -- -- -- 72.2 71.7 [72] Inner Retinoschisis -- -- 79.4% -- -- -- 72.2 72.3 Outer Retinoschisis -- -- 79.4% -- -- -- 72.2 80.1 Retinal Detachment -- -- 79.4% -- -- -- 72.2 85.1 Du et al., 2022 (unspecified backbone) 9176 OCT images Myopic Neovascularizatoin 0.985 -- -- -- -- -- -- -- [74] Myopic Traction Maculopathy 0.946 -- -- -- -- -- -- -- Dome Shaped Macula 0.978 -- -- -- -- -- -- -- Chen et al., 2023 Multi-System (ResNet50, YOLOv3, Ensemble CNNs) 37,138 OCT images Vitreomacular Traction Syndrome -- 99.3% 98.4% 99.5% 98.4 -- -- -- [ 73 ] Macular Pucker -- -- 98.2% -- -- 98.2% -- -- Cystoid Macular Edema -- -- 97.7% -- -- 93.5% -- -- Full Thickness Retinal Eminence -- -- 95.9% -- -- 92.2% -- -- Epiretinal Membrane -- -- 87.0% -- -- 67.0% -- -- Open in a new tab AUC : area under curve; IOU : intersection over union; DSC : dice similarity coefficient; OCT : optical coherence tomography For instance, Huang et al., implemented a classification algorithm to demonstrate the utility of the ResNet-34 architecture in the staging of MTM from OCT images [ 70 ]. They generated an independent five-class schema that categorized an image as having either no MTM (class 0), extra-foveal maculoschisis (class 1), inner lamellar macular hole (class 2), outer foveoschisis (class 3), or foveal detachment (class 4). Their proposed model performed exceedingly well, achieving an AUC of 0.993, 0.996, 0.984, 0.996, and 0.998 and accuracy of 0.975, 0.977, 0.982, 0.992, and 0.992 in the diagnosis of each of the respective five classes. Furthermore, its performance was found to be equivalent to or even better than retina specialists when classifying according to this classification schema. However, the algorithm constructed by Huang et al., is incapable of localizing retinal lesions and was also designed strictly for the classification of MTM but not other pathologies associated with high myopia, such as Bruch’s membrane defects, choroidal neovascularization, and macular choroidal thinning. Its design as a single-institution study and relatively smaller test set of 604 OCT images also limit it generalizability. In contrast, Chen et al. trained models built off the ResNet-50 and a YOLOv3 architecture to diagnose not only vitreomacular traction syndrome (VMT) but also AMD, along with ten specific subtypes of lesions associated with both pathologies, from OCT images [ 71 ]. Although VMT and MTM are distinct entities, this study is relevant because it demonstrates OCT-based tractional pathology detection and lesion localization methods that are conceptually transferable to MTM-focused datasets.Their multi-stage classification system first categorized images as normal and abnormal with an ensemble of CNNs. Resnet-50 was then applied to diagnose abnormal images as either VMT or AMD. When doing so, their model was able to detect VMT with an accuracy, sensitivity, and specificity of 99.3%, 98.4%, and 99.5%, respectively. However, by integrating a YOLOv3 architecture and training their model with OCT images with retinal lesions labeled by ophthalmologists, the model designed by Chen et al. surpasses that of Huang et al., in its ability to localize specific defects associated with VMT and AMD, such as macular pucker, cystoid macular edema, epiretinal membranes, hemorrhages, atrophy, and others. In doing so, their model exhibited better performance when detecting VMT lesions, demonstrating a recall and precision ranging between 87.0 and 98.2% and 67.0–98.2% across the different lesion subtypes, respectively. Additionally, though their study was similarly derived from a single institution, the test set of Chen et al., utilizes 3,126 OCT images, yielding more generalizable data than Huang et al. Taken together, these classification models generated by Huang et al., and Chen et al., illustrate the substantial clinical potential of DL-based approaches to precisely identify and differentiate MTM subtypes, ultimately enabling enhanced diagnostic accuracy, informed therapeutic decisions, and improved patient outcomes. Segmentation models in the detection of MTM Although the distinct visual features of MTM are well-suited for direct classification systems, segmentation-assisted models may offer additional advantages and improved diagnostic precision. In Chen et al.’s study, after classifying OCT images as VMT, the images were subsequently analyzed using the YOLOv3 deep learning architecture for lesion localization [ 71 ]. Its overall accuracy, sensitivity, and specificity for the detection of VMT was reported to be 99.3%, 98.4%, and 99.5%, respectively [ 71 ]. However, the employment of YOLOv3 integrated an object detection algorithm to identify specific features within images, providing detailed clinical insights during OCT scan analysis. For example, when programmed to detect lesions associated with macular pucker, a complication of VMT and feature of MTM, the model did so with a recall of 98.2% and precision of 98.2%. Meanwhile, when localizing other VMT lesions such as cystoid macular edema, full-thickness retinal eminences, epiretinal membranes, and regions of increased retinal thickness by detachments, the model yielded a recall of 97.7%, 95.9%, 87.0%, and 90.9% and a precision of 93.5%, 92.2%, 67.0%, and 74.1%, respectively. However, despite the performance in Chen et al. their YOLOv3-based system performs object classification by labeling lesion locations with rectangular boxes. This approach localizes lesions but does not delineate their boundaries at the pixel-level, which is characteristic across true segmentation architectures. By contrast, Jiang et al., combined CNNs with transformers to develop the HyFormer model, designed specifically for lesion segmentation in MTM OCT images (Table 3 ) [ 72 ]. Running images through these two systems in parallel allows HyFormer to capture both the global and minute details associated with the images of different lesions. This hybrid approach enabled accurate identification and segmentation of internal limiting membrane detachments (ILMD), inner retinoschisis (IRS), outer retinoschisis (ORS), and retinal detachments (RD) in MTM patients at the pixel-level. The system created by Jiang et al. displayed a DSC of 71.72%, 72.29%, 80.06%, and 85.13%, respectively, in the segmentation of ILMD, IRS, ORS, and RD, with an average IOU of 72.17%. Thus, the HyFormer model surpasses that of Chen et al. in its ability to not only localize lesions but also delineate their boundaries at the pixel-level. Importantly, the segmentation performance of the HyFormer model surpassed that of most other contemporary DL models across all four categories as well, underscoring the potential of hybrid segmentation models in enhancing the interpretation of OCT scans in MTM diagnosis [ 72 ]. The unmet needs and implications of AI in the detection of MTM DL models have demonstrated effectiveness in diagnosing MTM from OCT images, yet their ability to accurately prognosticate disease progression remains constrained. This limitation is particularly concerning given the progressive nature of MTM and its distinct subtypes. Horizontal (tangential) and perpendicular MTM, distinguished by the direction of tractional pull, differ significantly in their clinical implications and prognosis [ 75 ]. For instance, tangential MTM, commonly associated with epiretinal membranes or internal limiting membrane abnormalities, is typically less severe and progresses more slowly. In contrast, perpendicular MTM, linked to posterior vitreous detachments and progressive staphylomas, carries a higher risk of rapid progression and retinal breaks [ 75 ]. These differences underscore the importance of longitudinally tracking the subtle structural changes associated with each subtype to accurately stratify patient risk and tailor interventions accordingly. Currently, prognostic insights for MTM are primarily derived from retrospective observational studies using conventional statistical analyses. For example, Li et al. documented influential predictors of MTM progression, such as outer retinoschisis location, partial posterior vitreous detachment, and disruption of the ellipsoid zone, affecting visual outcomes [ 76 ]. However, these traditional models do not utilize the advanced pattern-recognition capabilities provided by AI. AI-driven predictive models could significantly enhance current prognostic methods by automatically integrating complex imaging biomarkers and longitudinal clinical data to more precisely monitor disease progression. Future AI models that can automatically detect subtle worsening of traction on OCT imaging over time could proactively predict retinal breaks or rapid progression, thereby informing earlier surgical decision-making, including pars plana vitrectomy or macular buckling [ 77 ]. Nonetheless, the advancement of such AI-based prognostic models depends heavily on the availability of detailed and robust longitudinal datasets, highlighting a critical area for future data collection and research efforts. For prognostic models intended to guide follow-up or timing of surgery, calibration and clinically meaningful risk horizons are as important as discrimination metrics. Similarly, surgical outcomes for MTM patients can vary widely, yet predictive models specifically targeting functional outcomes post-operatively remain underdeveloped. While existing predictive studies have identified important anatomical features influencing outcomes, such as baseline visual acuity and the presence of foveal retinal detachment, current models primarily focus on anatomical endpoints rather than functional visual results [ 78 ]. Consequently, these models may determine anatomical success after vitrectomy, but are less capable of predicting how a patient’s visual acuity will change post-operatively. This limitation may result from a lack of multi-modal AI approaches. Advancements in AI technologies that simultaneously integrate clinical data from OCT, fundus photography, fluorescein angiography, and other imaging modalities could enhance predictive accuracy. Training AI models to correlate detailed image interpretations with functional visual outcomes in the postoperative period could further refine their prognostic capabilities. Ultimately, expanding AI models in this manner would inform surgical decision-making, help manage patient expectations preoperatively, and better identify patients most likely to benefit from surgical interventions. Conclusions The surging global prevalence of myopia warrants novel approaches towards the management of pathologic myopia [ 1 ]. Integration of AI models into ophthalmic practice offers innovative diagnostic and prognostic solutions to address this increasing clinical burden. These models have demonstrated promise in the diagnosis of pathologic myopia, along with myopic macular degeneration, myopic traction maculopathy, and retinal detachments. By applying classification and segmentation algorithms, DL models can identify retinal lesions from the outputs of fundus photography, OCT scans, and B-scan ultrasonography. Supplementing ophthalmologists’ clinical decision making with this technology can encourage more timely diagnoses and interventions and lead to better visual outcomes in patients with high myopia. More so, it allows for more streamlined practice by reducing burdens on clinic staff since these models can assist with triage and preliminary image interpretation that supports decisions made by clinicians and patients. By supporting triage and remote interpretation in settings with limited specialist availability, AI-driven medicine may promote more equitable access to care. Disadvantaged communities that lack ophthalmologists can have images interpreted remotely with specialists rather than backlogging referrals to eye clinics. Thus, integrating AI into clinical ophthalmology has the potential to significantly improve patient care, streamline clinical workflows, and expand access to specialized ophthalmic services globally. However, despite their effectiveness in diagnostics, more work is necessary to strengthen the ability of deep learning models to properly prognosticate the retinal complications of pathologic myopia. Doing so can expand the clinical implications of this technology towards predicting patients’ surgical outcomes, gauging longitudinal visual acuities, identifying high risk patients in need of timely interventions, and personalizing treatment regimens. The application of longitudinal and diverse training databases can serve as a potential resolution, empowering physicians to fully harness the potential of AI. Training deep learning models with databases containing images from different modalities and institutions can expand their knowledge and bandwidth. Larger and more diverse training data can improve reliability and reduce spurious correlations, but interpretability and clinical trust also require transparent reporting, calibration, and clinically meaningful explanation methods. This can have large implications in workspaces where there may be hesitancy to rely on AI for clinical decisions, encouraging eye care providers to collaborate with this technology. Additionally, diversifying the training of AI can make this technology more generalizable to a larger number of communities. By integrating the clinical data over diverse communities, the outputs of these models can be applicable to more patients, ultimately culminating in more equitable eye care. Supplementary Information Below is the link to the electronic supplementary material. Supplementary figure 1 (2.6MB, png) Supplementary File 1.Optos fundus photography image of an OD bullous rhegmatogenous macula-off, fovea-offretinal detachment (PNG 2.64 MB) High Resolution Image (TIFF 4.62 MB) (4.6MB, tiff) Supplementary figure 2 (2MB, png) Supplementary File 2.Fundus autofluorescence photography (left) and optical coherence tomography (right)images of an OD stage 3a (international classification) myopic traction maculopathy (PNG 1.95 MB) High Resolution Image (TIFF 2.91 MB) (2.9MB, tiff) Supplementary figure 3 (3.8MB, png) Supplementary File 3.Optos fundus photography (top left), fundus autofluorescence photography (top right), andoptical coherence tomography (bottom) images of myopic macular atrophy complicated by choroidalneovascularization (PNG 3.83 MB) High Resolution Image (TIFF 5.96 MB) (6MB, tiff) Author contributions Conceptualization : Yeabsira Mesfin, Chase A. Ludwig; Literature Search : Yeabsira Mesfin, Chase A. Ludwig; Data Analysis : Yeabsira Mesfin; Writing - original draft : Yeabsira Mesfin; Writing - review and editing : Yeabsira Mesfin, Anish Salvi, Leo Arnal, Curtis Langlotz, Vinit Mahajan, Chase A. Ludwig; Funding acquisition : Chase A. Ludwig. Funding This work was supported by the National Eye Institute K23 Grant, K23EY035741 and E. Matilda Ziegler Foundation for the Blind Grant awarded to Chase A. Ludwig as well as the Stanford P30 Vision Research Core Grant, NEI P30-EY026877, and Research to Prevent Blindness, Inc for Chase A. Ludwig. Declarations Ethical approval All procedures performed in studies involving human participants were in accordance with the ethical standards of Stanford University and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Informed consent The use of anonymized, aggregated patient data rendered this study exempt from requiring patient consent. Conflicts of interest All authors certify that they have no affiliations with or involvement in any organization or entity with any financial interest (such as honoraria; educational grants; participation in speakers’ bureaus; membership, employment, consultancies, stock ownership, or other equity interest; and expert testimony or patent-licensing arrangements), or non-financial interest (such as personal or professional relationships, affiliations, knowledge or beliefs) in the subject matter or materials discussed in this manuscript. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Holden BA, Fricke TR, Wilson DA et al (2016) Global prevalence of myopia and high myopia and temporal trends from 2000 through 2050. Ophthalmology 123(5):1036–1042 [ DOI ] [ PubMed ] [ Google Scholar ] 2. Jones LA, Sinnott LT, Mutti DO, Mitchell GL, Moeschberger ML, Zadnik K (2007) Parental history of myopia, sports and outdoor activities, and future myopia. Invest Ophthalmol Vis Sci 48(8):3524–3532 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. McCrann S, Loughman J, Butler JS, Paudel N, Flitcroft DI (2021) Smartphone use as a possible risk factor for myopia. Clin Exp Optom 104(1):35–41 [ DOI ] [ PubMed ] [ Google Scholar ] 4. Hepsen IF, Evereklioglu C, Bayramlar H (2001) The effect of reading and near-work on the development of myopia in emmetropic boys: a prospective, controlled, three-year follow-up study. Vis Res 41(19):2511–2520 [ DOI ] [ PubMed ] [ Google Scholar ] 5. Ha A, Lee YJ, Lee M, Shim SR, Kim YK (2025) Digital screen time and myopia: a systematic review and dose-response meta-analysis. JAMA Netw Open 8(2):e2460026–e2460026 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Quiroz-Reyes MA, Quiroz-Gonzalez EA, Quiroz-Gonzalez MA, Lima-Gomez V (2024) Comprehensive assessment of glaucoma in patients with high myopia: a systematic review and meta-analysis with a discussion of structural and functional imaging modalities. Int Ophthalmol 44(1):405 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Pan C-W, Cheng C-y, Saw S-M, Wang JJ, Wong TY (2013) Myopia and age-related cataract: a systematic review and meta-analysis. Am J Ophthalmol 156(5):1021-1033. e1 [ DOI ] [ PubMed ] [ Google Scholar ] 8. Han X, Ong J-S, An J et al (2020) Association of myopia and intraocular pressure with retinal detachment in European descent participants of the UK biobank cohort: a Mendelian randomization study. JAMA Ophthalmol 138(6):671–678 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Jonas JB, Ohno-Matsui K, Panda-Jonas S (2019) Myopia: anatomic changes and consequences for its etiology. Asia-Pac J Ophthalmol 8(5):355–359 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Naidoo KS, Fricke TR, Frick KD et al (2019) Potential lost productivity resulting from the global burden of myopia: systematic review, meta-analysis, and modeling. Ophthalmology 126(3):338–346 [ DOI ] [ PubMed ] [ Google Scholar ] 11. Teuwen J, Moriakov N (2020) Convolutional neural networks. Handbook of medical image computing and computer assisted intervention. Elsevier, pp 481–501 12. Liu X, Gao K, Liu B et al (2021) Advances in deep learning-based medical image analysis. Health Data Sci 2021:8786793 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Wu H, Liu Q, Liu X (2019) A review on deep learning approaches to image classification and object segmentation. Comput Mater Contin. 10.32604/cmc.2019.03595 [ Google Scholar ] 14. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. :770–778 15. Gouda N, Amudha J (2020) Skin cancer classification using ResNet. IEEE, pp 536–541 16. Zhuang Q, Gan S, Zhang L (2022) Human-computer interaction based health diagnostics using ResNet34 for tongue image classification. Comput Methods Programs Biomed 226:107096 [ DOI ] [ PubMed ] [ Google Scholar ] 17. Subaar C, Addai FT, Addison ECK et al (2024) Investigating the detection of breast cancer with deep transfer learning using ResNet18 and ResNet34. Biomedical Physics & Engineering Express 10(3):035029 [ DOI ] [ PubMed ] [ Google Scholar ] 18. Asadi-Aghbolaghi M, Darbandsari A, Zhang A et al (2024) Learning generalizable AI models for multi-center histopathology image classification. NPJ Precis Oncol 8(1):151 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Cho J, Lee K, Shin E, Choy G, Do S (2015) How much data is needed to train a medical image deep learning system to achieve necessary high accuracy? arXiv preprint arXiv:151106348 20. Rocha A, Papa JP, Meira LA (2012) How far do we get using machine learning black-boxes? Int J Pattern Recognit Artif Intell 26(02):1261001 [ Google Scholar ] 21. Tamori H, Yamashina H, Mukai M, Morii Y, Suzuki T, Ogasawara K (2022) Acceptance of the use of artificial intelligence in medicine among Japan’s doctors and the public: a questionnaire survey. JMIR Hum Factors 9(1):e24680 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Arun N, Gaw N, Singh P et al (2021) Assessing the trustworthiness of saliency maps for localizing abnormalities in medical imaging. Radiology: Artif Intell 3(6):e200267 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Minaee S, Boykov Y, Porikli F, Plaza A, Kehtarnavaz N, Terzopoulos D (2021) Image segmentation using deep learning: a survey. IEEE Trans Pattern Anal Mach Intell 44(7):3523–3542 [ DOI ] [ PubMed ] [ Google Scholar ] 24. Moghbel M, Mashohor S, Mahmud R, Saripan MIB (2018) Review of liver segmentation and computer assisted detection/diagnosis methods in computed tomography. Artif Intell Rev 50:497–537 [ Google Scholar ] 25. Rajinikanth V, Kadry S, Nam Y (2021) Convolutional-neural-network assisted segmentation and SVM classification of brain tumor in clinical MRI slices. Inf Technol Control 50(2):342–356 [ Google Scholar ] 26. Blankemeier L, Cohen JP, Kumar A et al (2024) Merlin: A vision language foundation model for 3d computed tomography. Research Square . :rs. 3. rs–4546309 27. Weese J, Lorenz C (2016) Four challenges in medical image analysis from an industrial perspective. Elsevier, pp 44–49 [ DOI ] [ PubMed ] 28. Steel D (2014) Retinal detachment. BMJ Clin Evid 2014:0710 [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Pierro L, Camesasca F, Mischi M, Brancato R (1992) Peripheral retinal changes and axial myopia. Retina 12(1):12–17 [ DOI ] [ PubMed ] [ Google Scholar ] 30. Lakawicz JM, Bottega WJ, Fine HF, Prenner JL (2020) On the mechanics of myopia and its influence on retinal detachment. Biomech Model Mechanobiol 19:603–620 [ DOI ] [ PubMed ] [ Google Scholar ] 31. Wang Y-h, Huang C, Tseng Y-l, Zhong J, Li X-m (2021) Refractive error and eye health: an umbrella review of meta-analyses. Front Med (Lausanne) 8:759767 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Zhang C, He F, Li B et al (2021) Development of a deep-learning system for detection of lattice degeneration, retinal breaks, and retinal detachment in tessellated eyes using ultra-wide-field fundus images: a pilot study. Graefes Arch Clin Exp Ophthalmol. 10.1007/s00417-021-05105-3 [ DOI ] [ PubMed ] [ Google Scholar ] 33. Ohsugi H, Tabuchi H, Enno H, Ishitobi N (2017) Accuracy of deep learning, a machine-learning technology, using ultra–wide-field fundus ophthalmoscopy for detecting rhegmatogenous retinal detachment. Sci Rep 7(1):9425 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Yadav S, Mandal S, Murugan R, Goel T, Ahmed T (2024) Segmentation and visualization of retinal detachment lesions through retinal fundus images. Biomed Signal Process Control 96:106627 [ Google Scholar ] 35. Caki O, Guleser UY, Ozkan D et al (2025) Automated detection of retinal detachment using deep Learning-Based segmentation on ocular ultrasonography images. Translational Vis Sci Technol 14(2):26–26 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Christ M, Habra O, Monnin K et al (2024) Deep learning-based automated detection of retinal breaks and detachments on fundus photography. Transl Vis Sci Technol 13(4):1–1 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Zhou W-D, Dong L, Zhang K et al (2022) Deep learning for automatic detection of recurrent retinal detachment after surgery using ultra-widefield fundus images: a single‐center study. Adv Intell Syst 4(9):2200067 [ Google Scholar ] 38. Li Z, Guo C, Nie D et al (2020) Deep learning for detecting retinal detachment and discerning macular status using ultra-widefield fundus images. Commun Biol 3(1):15 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Rashid MM (2025) Detection of Retinal Detachment Using Deep Learning and Data Mining Approaches 40. Chen D, Yu Y, Zhou Y et al (2021) A deep learning model for screening multiple abnormal findings in ophthalmic ultrasonography (with video). Transl Vis Sci Technol 10(4):22–22 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Younis YB, Al Azzo F (2025) A technique for retinal detachment detection manipulating YOLOv8 models. NTU J Eng Technol. ;4(1) 42. Ng DS, Chan LK, Lai TY (2023) Myopic macular diseases: a review. Clin Exp Ophthalmol 51(3):229–242 [ DOI ] [ PubMed ] [ Google Scholar ] 43. Ruiz-Medrano J, Montero JA, Flores-Moreno I, Arias L, García-Layana A, Ruiz-Moreno JM (2019) Myopic maculopathy: current status and proposal for a new classification and grading system (ATN). Prog Retin Eye Res 69:80–115 [ DOI ] [ PubMed ] [ Google Scholar ] 44. Yokoi T, Ohno-Matsui K (2018) Diagnosis and treatment of myopic maculopathy. Asia-Pac J Ophthalmol 7(6):415–421 [ DOI ] [ PubMed ] [ Google Scholar ] 45. Ohno-Matsui K, Kawasaki R, Jonas JB et al (2015) International photographic classification and grading system for myopic maculopathy. Am J Ophthalmol 159(5):877–883 e7 [ DOI ] [ PubMed ] [ Google Scholar ] 46. Wong Y-L, Sabanayagam C, Ding Y et al (2018) Prevalence, risk factors, and impact of myopic macular degeneration on visual impairment and functioning among adults in Singapore. Investig Ophthalmol Vis Sci 59(11):4603–4613 [ DOI ] [ PubMed ] [ Google Scholar ] 47. Ohno-Matsui K, Wu P-C, Yamashiro K et al (2021) IMI pathologic myopia. Invest Ophthalmol Vis Sci 62(5):5–5 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Flitcroft DI, He M, Jonas JB et al (2019) IMI–defining and classifying myopia: a proposed set of standards for clinical and epidemiologic studies. Investig Ophthalmol Vis Sci 60(3):M20–M30 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Wu Z, Cai W, Xie H et al (2022) Predicting optical coherence tomography-derived high myopia grades from fundus photographs using deep learning. Front Med 9:842680 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Zhao J, Cao G, He J, Dai C (2024) Multi-class classification of pathological myopia based on fundus photography. J Innov Opt Health Sci 17(6):2450016-2450016–12 [ Google Scholar ] 51. Tang J, Yuan M, Tian K et al (2022) An artificial-intelligence–based automated grading and lesions segmentation system for myopic maculopathy based on color fundus photographs. Transl Vis Sci Technol 11(6):16–16 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Ali S, Raut S (2024) Smartphone app to detect pathological myopia using spatial attention and squeeze-excitation network as a classifier and segmentation encoder. Int J Imaging Syst Technol 34(5):e23157 [ Google Scholar ] 53. Du R, Xie S, Fang Y et al (2021) Deep learning approach for automated detection of myopic maculopathy and pathologic myopia in fundus images. Ophthalmol Retina 5(12):1235–1244 [ DOI ] [ PubMed ] [ Google Scholar ] 54. Lu L, Ren P, Tang X et al (2021) AI-model for identifying pathologic myopia based on deep learning algorithms of myopic maculopathy classification and “plus” lesion detection in fundus images. Front Cell Dev Biol 9:719262 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Park S-J, Ko T, Park C-K, Kim Y-C, Choi I-Y (2022) Deep learning model based on 3D optical coherence tomography images for the automated detection of pathologic myopia. Diagnostics 12(3):742 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Choi KJ, Choi JE, Roh HC et al (2021) Deep learning models for screening of high myopia using optical coherence tomography. Sci Rep 11(1):21663 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 57. Ye X, Wang J, Chen Y et al (2021) Automatic screening and identifying myopic maculopathy on optical coherence tomography images using deep learning. Translational Vis Sci Technol 10(13):10–10 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Rauf N, Gilani SO, Waris A (2021) Automatic detection of pathological myopia using machine learning. Sci Rep 11(1):16570 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 59. Wang R, He J, Chen Q et al (2023) Efficacy of a deep learning system for screening myopic maculopathy based on color fundus photographs. Ophthalmol Therapy 12(1):469–484 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Hemelings R, Elen B, Blaschko MB, Jacob J, Stalmans I, De Boever P (2021) Pathological myopia classification with simultaneous lesion segmentation using deep learning. Comput Methods Programs Biomed 199:105920 [ DOI ] [ PubMed ] [ Google Scholar ] 61. Chen Y, Yang S, Liu R et al (2024) Forecasting myopic maculopathy risk over a decade: development and validation of an interpretable machine learning algorithm. Investig Ophthalmol Vis Sci 65(6):40–40 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Foo LL, Xu L, Sabanayagam C et al (2023) Predictors of myopic macular degeneration in a 12-year longitudinal study of Singapore adults with myopia. Br J Ophthalmol 107(9):1363–1368 [ DOI ] [ PubMed ] [ Google Scholar ] 63. Choudhary A, Venkatesh RH, Jayashree M, Surendrappa HD, Divya R, Darshini L (2021) Effect of high myopia on macular thickness: an optical coherence tomography study in a tertiary care hospital, Karnataka, India. J Clin Ophthalmol Res 9(1):14–17 [ Google Scholar ] 64. Sayah DN, Garg I, Katz R et al (2023) Characterizing macular neovascularization in myopic macular degeneration and age-related macular degeneration using swept source OCTA. Clin Ophthalmol. 10.2147/OPTH.S440575 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 65. Chen Y, Han X, Gordon I et al (2022) A systematic review of clinical practice guidelines for myopic macular degeneration. J Glob Health 12:04026 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 66. Chierigo A, Ferro Desideri L, Traverso CE, Vagge A (2022) The role of atropine in preventing myopia progression: an update. Pharmaceutics 14(5):900 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 67. Panozzo G, Mercanti A (2004) Optical coherence tomography findings in myopic traction maculopathy. Arch Ophthalmol 122(10):1455–1460 [ DOI ] [ PubMed ] [ Google Scholar ] 68. Ouyang P-B, Duan X-C, Zhu X-H (2012) Diagnosis and treatment of myopic traction maculopathy. Int J Ophthalmol 5(6):754 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 69. Raizada K, Sahu J (2020) Myopic Foveoschisis [ PubMed ] 70. Huang X, He S, Wang J, Yang S, Wang Y, Ye X (2023) Lesion detection with fine-grained image categorization for myopic traction maculopathy (MTM) using optical coherence tomography. Med Phys 50(9):5398–5409 [ DOI ] [ PubMed ] [ Google Scholar ] 71. Chen X, Xue Y, Wu X et al (2023) Deep learning-based system for disease screening and pathologic region detection from optical coherence tomography images. Transl Vis Sci Technol 12(1):29–29 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 72. Jiang Q, Fan Y, Li M et al (2024) HyFormer: a hybrid transformer-CNN architecture for retinal OCT image segmentation. Biomed Opt Express 15(11):6156–6170 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 73. Chen S, Wu Z, Li M et al (2023) Fit-net: feature interaction transformer network for pathologic myopia diagnosis. IEEE Trans Med Imaging 42(9):2524–2538 [ DOI ] [ PubMed ] [ Google Scholar ] 74. Du R, Xie S, Fang Y et al (2022) Validation of soft labels in developing deep learning algorithms for detecting lesions of myopic maculopathy from optical coherence tomographic images. Asia-Pacific J Ophthalmol 11(3):227–236 [ DOI ] [ PubMed ] [ Google Scholar ] 75. Anderson WJ, Akduman L (2023) Management of myopic maculopathy: a review. Turk J Ophthalmol 53(5):307 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. Li S, Li T, Wang X et al (2021) Natural course of myopic traction maculopathy and factors influencing progression and visual acuity. BMC Ophthalmol 21:1–11 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 77. Parolini B, Palmieri M, Finzi A, Frisina R (2021) Proposal for the management of myopic traction maculopathy based on the new MTM staging system. Eur J Ophthalmol 31(6):3265–3276 [ DOI ] [ PubMed ] [ Google Scholar ] 78. Hattori K, Kataoka K, Takeuchi J, Ito Y, Terasaki H (2018) Predictive factors of surgical outcomes in vitrectomy for myopic traction maculopathy. Retina 38:S23–S30 [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary figure 1 (2.6MB, png) Supplementary File 1.Optos fundus photography image of an OD bullous rhegmatogenous macula-off, fovea-offretinal detachment (PNG 2.64 MB) High Resolution Image (TIFF 4.62 MB) (4.6MB, tiff) Supplementary figure 2 (2MB, png) Supplementary File 2.Fundus autofluorescence photography (left) and optical coherence tomography (right)images of an OD stage 3a (international classification) myopic traction maculopathy (PNG 1.95 MB) High Resolution Image (TIFF 2.91 MB) (2.9MB, tiff) Supplementary figure 3 (3.8MB, png) Supplementary File 3.Optos fundus photography (top left), fundus autofluorescence photography (top right), andoptical coherence tomography (bottom) images of myopic macular atrophy complicated by choroidalneovascularization (PNG 3.83 MB) High Resolution Image (TIFF 5.96 MB) (6MB, tiff) Articles from Graefe's Archive for Clinical and Experimental Ophthalmology are provided here courtesy of Springer ACTIONS View on publisher site PDF (1.3 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top