Imaging-based machine learning for the diagnosis and prognosis of uveal melanoma: a systematic review and meta analysis.

backgroundUveal melanoma (UM) is the most common primary intraocular malignancy in adults and carries a high risk of metastasis and poor prognosis when diagnosed late. Distinguishing UM from benign choroidal nevi remains challenging due to overlapping imaging features. Machine learning (ML) and deep learning (DL) models have emerged as tools to improve diagnostic accuracy and prognostic prediction, but their generalizability and clinical readiness remain uncertain. PURPOSE: To systematically review and meta-analyze the performance of published ML/DL algorithms using ocular imaging on detecting and predicting prognosis of UM.

methodsA systematic search was conducted in PubMed, Scopus, Web of Science, Embase, and IEEE Xplore up to June 2025 for English and Spanish publications since 2012. Eligible studies applied ML/DL to ocular imaging and reported diagnostic or prognostic accuracy metrics. Risk of bias was assessed with QUADAS-2, and pooled sensitivity and specificity were estimated using a random-effects meta-analysis, with subgroup analyses by imaging modality (fundus-only vs. ultrasound-based imaging, with or without an additional modality). For diagnostic studies, we additionally synthesized positive and negative likelihood ratios along with diagnostic odds ratios (DOR) and conducted hierarchical summary receiver operating characteristic curve (ROC) (Reitsma) analyses stratified by modality.

resultsThirteen diagnostic studies met inclusion criteria. Most were retrospective, single-center cohorts using fundus photography, while others employed ultrasound (US), ultra-widefield (UWF), OCT, or autofluorescence. Ten studies used DL architectures, mainly convolutional neural networks or transformers. Pooled sensitivity was 78.21% and specificity 94.28%. Fundus-based models showed lower specificity (87.70%) than US-based models (98.16%). AUC values were consistently high; HSROC AUC was 0.912 (fundus) vs. 0.984 (US; Δ0.073, bootstrap 95% CI − 0.101 to 0.143). The largest modality separation was for LR+ (39.48 vs. 6.16; p < 0.0001), with higher DOR for US (225.25 vs. 27.89; p = 0.0102), while LR − was similar. Six prognostic studies (up to 4,600 patients) reported AUCs between 0.71 and 0.92 for predicting metastasis, survival, or enucleation, though all lacked external validation.

conclusionsML and DL models show strong diagnostic performance and emerging prognostic value in UM. However, reproducibility and real-world validation remain limited. New foundation models such as RETFound and VisionFM, trained on large multimodal eye datasets, could improve standardization, explainability, and cross-center generalization. As these models evolve, they have the potential to become essential tools in clinical practice, accelerating the translation of AI into reliable, routine implementation in ocular oncology. (PROSPERO ID: CRD42025643874).