AI in the Shadows: Unveiling the Strengths and Blind Spots of Medios AI Retinal Screening in Cancer Care

Introduction Artificial intelligence (AI) is increasingly being integrated into ophthalmic diagnostics, offering potential for efficient screening of common retinal diseases. The Medios AI system by Remidio (Singapore, Singapore), designed for use with a smartphone-based fundus camera, claims to detect diabetic retinopathy (DR), age-related macular degeneration (ARMD), and glaucoma. However, its performance in complex clinical settings such as ocular oncology remains underexplored. This study aims to evaluate both the diagnostic capabilities and limitations of the Medios AI system when applied to a diverse cohort of oncology patients. Materials and methods An observational study was conducted over three months in an ocular oncology clinic. Ninety-eight cancer patients (196 eyes) underwent fundus photography using the Remidio smartphone-based fundus imaging system. The images were analyzed using the Medios AI algorithm. AI-generated findings were compared with clinical evaluations by an experienced ophthalmologist to identify diagnostic concordance and discrepancies. Additional attention was paid to the system's imaging capabilities, including its ability to capture wide-field or montage images. Result The AI system accurately identified glaucomatous cupping in three patients, flagged two cases of DR, and detected signs of ARMD in two patients-all consistent with clinical examination. However, eight patients with leukemic retinopathy were incorrectly flagged as having DR, revealing a lack of specificity in distinguishing vascular retinal pathologies. The system also failed to detect optic atrophy, a critical neuro-ophthalmic finding in oncology patients. A technical limitation was also noted: the inability of the Remidio system to generate montage or wide-field images, restricting visualization of the peripheral retina. Conclusion While the Medios AI system demonstrates promise in identifying common retinal pathologies such as DR, ARMD, and glaucoma, its limitations are significant in the oncology context. The inability to distinguish similar hemorrhagic retinopathies, failure to detect optic nerve atrophy, and lack of wide-field imaging capabilities underscore the need for cautious implementation. Integration of AI tools must be accompanied by expert clinical oversight, especially in specialized settings where retinal presentations are complex and atypical.