Geographic atrophy (GA) is a progressive retinal disease associated with late-stage age-related macular degeneration (AMD), a significant cause of visual impairment in senior adults. GA lesion segmentation is important for disease monitoring in clinical trials and routine ophthalmic practice; however, its manual delineation is time-consuming, laborious, and subject to inter-grader variability. The use of artificial intelligence (AI) is rapidly expanding within the medical field and could potentially improve accuracy while reducing the workload by facilitating this task. This systematic review evaluates the performance of AI algorithms for GA segmentation and highlights their key limitations from the literature. Five databases and two registries were searched from inception until 23 March 2024, following the PRISMA methodology. Twenty-four studies met the prespecified eligibility criteria, and fifteen were included in this meta-analysis. The pooled Dice similarity coefficient (DSC) was 0.91 (95% CI 0.88–0.95), signifying a high agreement between the reference standards and model predictions. The risk of bias and reporting quality were assessed using QUADAS-2 and CLAIM tools. This review provides a comprehensive evaluation of AI applications for GA segmentation and identifies areas for improvement. The findings support the potential of AI to enhance clinical workflows and highlight pathways for improved future models that could bridge the gap between research settings and real-world clinical practice.

This review examines the role of artificial intelligence (AI) in differentiating choroidal nevus (CN) from choroidal melanoma (CM), focusing on diagnosis, classification, segmentation, and prediction of malignant transformation. A literature search was performed in PubMed, Google Scholar, and Science Direct up to December 2024. Eight studies met all the inclusion criteria, evaluating machine learning (ML) and deep learning (DL) applications for CN and CM assessment using various ophthalmic imaging modalities. Performance varied across AI models. U-Net achieved 100% sensitivity and an AUC of 0.88, while DenseNet121 reached an AUC of 0.9781, LASSO logistic regression an AUC of 0.88, RETFound (a self-supervised learning model) had an AUCROC of 0.92, and ResNet50 an accuracy of 92.65% in classification tasks. DeepLabv3 achieved Dice scores of 0.87 (melanoma) and 0.81 (nevi) in lesion-based segmentation, while nnU-Net yielded a Dice score of 0.78 and a recall of 0.77 for pigmented lesion segmentation. SAINT (XGBoost-based) demonstrated a strong predictive performance (AUC: 0.910), confirming its effectiveness in ophthalmic imaging. These results highlight the potential of AI models as effective diagnostic tools in ocular oncology. However, further research is needed to improve model generalizability, enhance dataset diversity, and facilitate clinical integration.

Epiretinal membrane (ERM) is a common retinal pathology associated with progressive visual impairment, requiring timely and accurate assessment. Recent advances in artificial intelligence (AI) have enabled automated approaches for ERM detection, segmentation, and postoperative best corrected visual acuity (BCVA) prediction, offering promising avenues to enhance clinical efficiency and diagnostic precision. We conducted a comprehensive literature search across MEDLINE (via PubMed), Scopus, CENTRAL, ClinicalTrials.gov, and Google Scholar from the inception to 31 December 2023. A total of 42 studies were included in the systematic review, with 16 eligible for meta-analysis. Risk of bias and reporting quality were assessed using the QUADAS-2 and CLAIM tools. Meta-analysis of 16 studies (533,674 images) showed that deep learning (DL) models achieved high diagnostic accuracy (AUC = 0.97), with pooled sensitivity and specificity of 0.93 and 0.97, respectively. Optical coherence tomography (OCT)-based models outperformed fundus-based ones, and although performance remained high under external validation, the positive predictive value (PPV) declined—highlighting the importance of testing model generalizability. To the best of our knowledge, this is the first systematic review and meta-analysis to critically evaluate the role of AI in the detection, segmentation, and postoperative BCVA prediction of ERM across various ophthalmic imaging modalities. Our findings provide a clear overview of current evidence supporting the continued development and clinical adoption of AI tools for ERM diagnosis and management.

This systematic review and meta‐analysis evaluates the performance of artificial intelligence (AI)‐based models for predicting the onset and progression of refractive error (RE) in children and adolescents and quantitatively synthesizes their prediction accuracy. The study was preregistered in the PROSPERO database (CRD420251075160) and conducted in accordance with PRISMA guidelines. MEDLINE (via PubMed), Web of Science and Scopus were searched without time restrictions. Eligible studies involved children and adolescents (≤18 years) and applied AI‐based technologies to predict the onset or progression of RE, using non‐image‐based input data. To evaluate the prediction accuracy of the included studies, a meta‐analysis was performed. Quality assessment was conducted using the PROBAST+AI tool. Sixteen studies were included in the analysis. All studies evaluated myopia, two additionally addressed hyperopia and none focused on astigmatism. Most models predicted continuous SE values across the refractive spectrum, whereas a subset focused on classification‐based outcomes, such as myopia onset or progression. Included studies used diverse input data, including age, gender and parental myopia. Eight studies comprising 18 independent evaluations were included in the meta‐analysis. The pooled root‐mean‐squared error (RMSE) for spherical equivalent prediction was 0.50 D (95% confidence intervals: 0.37–0.62). Between‐study heterogeneity was substantial ( τ ² = 0.0695, I ² = 100%), indicating major differences across models. This high heterogeneity reflects substantial differences in study populations, prediction targets and horizons, model architectures and validation strategies across included studies, and therefore, the pooled estimate should be interpreted as an average across diverse settings rather than a directly generalizable performance measure. Leave‐one‐out sensitivity analyses demonstrated that no single study materially influenced the pooled effect. Across all iterations, RMSE estimates ranged narrowly from 0.48 to 0.52 D, with confidence intervals overlapping the main analysis. The studies' quality assessment showed mostly low concern, apart from the analysis domain, in which 43.8% of studies were rated high risk. AI models using non‐image‐based clinical data demonstrate moderate accuracy for predicting paediatric RE onset and progression. However, substantial methodological heterogeneity and limited external validation indicate that standardized development and validation frameworks are required before clinical implementation.

This study presents a two-phase framework for selecting clinically plausible explainable artificial intelligence (XAI) heatmaps for retinal optical coherence tomography (OCT) classification. A six-class Swin Transformer model was trained and validated using a combined dataset consisting of a subset of the public OCT-C8 dataset and private data from a Greek tertiary hospital and externally evaluated on an independent dataset from a private ophthalmological institute. Diagnostic performance was high, achieving 97% accuracy in cross-validation and 91.82% on external evaluation. In Phase 1, one ophthalmologist and one artificial intelligence (AI) specialist independently assessed 100 heatmaps per method based on visual quality and anatomical plausibility, reducing the candidate methods to three. In Phase 2, 21 specialists evaluated the selected methods across multiple cases using a five-point Likert scale reflecting agreement between highlighted regions and the model diagnosis. The proposed Token contRAST map (TRAST) achieved the highest ratings, followed by Gradient-weighted Class Activation Mapping (Grad-CAM++), while Cosine-Grad Fusion Map (CGFM) showed the lowest performance. These findings reflect clinical plausibility rather than direct model interpretability and indicate that effective XAI in OCT imaging requires not only technical performance but also structured expert evaluation. The proposed framework provides a practical approach for selecting explanation methods suitable for clinical use in ophthalmology.

Purpose: To develop and validate a self‐supervised vision transformer (ViT) for automated three‐class classification of Normal, High Myopia (HM), and Pathologic Myopia (PM) from colour fundus photographs, with anatomically faithful interpretability. Methods: A DINOv2 self‐supervised ViT was fine‐tuned using a two‐stage transfer‐learning protocol with class‐balanced sampling. Model robustness was assessed using 10‐fold stratified cross‐validation, and an ensemble of the best fold‐specific checkpoints was evaluated once on a strictly held‐out independent test set. Comparative performance was assessed against ResNet‐50, VGG‐16, and EfficientNet‐B3 using paired, two‐sided Student's t ‐tests on fold‐level macro–F1 scores. Model explainability was examined using a Dual‐Rollout Attention approach that fuses early‐layer spatial detail with late‐layer semantic information. Results: On the independent test set, the ensemble achieved an overall accuracy of 97.03% (95% CI 96.75–97.32%) and a macro F1‐score of 97.08% (95% CI 96.80–97.36%). The DINOv2‐based framework outperformed ResNet‐50, VGG‐16, and EfficientNet‐B3 architectures (all p < 0.001). Dual‐Rollout Attention highlighted clinically relevant regions including optic disc morphology, peripapillary atrophy, and areas consistent with chorioretinal thinning. Conclusion: A self‐supervised ViT combined with transformer‐specific interpretability enables accurate, robust, and transparent automated classification of high and pathologic myopia from fundus photographs. This approach may support standardized assessment and scalable screening, pending external and prospective validation.