Deep learning for early diagnosis of uveal melanoma: a systematic review and meta-analysis

Abstract Background Uveal melanoma (UM) is a rare cancer with an estimated annual incidence of 6 incidences per million people. About half of UM patients develop distant metastases, mainly to the liver. After metastasis, prognosis is poor, with median survival under 1 year and limited treatment options. Thus, earlier diagnostic methods for UM are a critical unmet need. The aim of this study was to evaluate the accuracy (sensitivity, specificity, and combined F1 score) of deep learning algorithms in the differential diagnosis of individuals with uveal melanoma. Methods We searched PubMed, Scopus, and Web of Science for studies comparing UM patients to healthy individuals or those with ocular nevi using AI diagnostic tools. AI performance metrics were extracted, with clinical or expert-based assessment as the reference standard. The study adhered to PRISMA guidelines. Results Five studies comprising 6388 patients (2981 UM, 2563 nevi, and 844 healthy) were included. The mean age of UM patients ranged from 58 to 63.2 years; for nevi, from 58 to 66 years. Pooled sensitivity was 89.0% (95% CI 88.6–89.5%) with no heterogeneity ( I 2 = 0.0%, p = 0.9242); individual sensitivities ranged from 82.4 to 100.0%. Pooled specificity was 84.9% (95% CI 73.7–91.9%) with significant heterogeneity ( I 2 = 72.3%, p = 0.006), ranging from 73.7 to 95.0%. The largest cohort had a specificity of 76.0% (95% CI 75.1–76.9%). Conclusions While fundus imaging is widely used in outpatient care, multimodal imaging remains limited to specialized clinics. Developing software to analyze fundus images could improve early, noninvasive UM detection and offer a cost-effective diagnostic tool.