Radiomics enables the extraction of high-throughput quantitative features from ophthalmic images, allowing the identification of subvisual information that is imperceptible to the human eye and offering a novel strategy for the precise diagnosis and treatment of retinal diseases. By quantitatively characterizing subtle differences in retinal structure, texture, and hemodynamic characteristics, and integrating these features with clinical data, radiomics has demonstrated substantial potential in early screening, disease stratification, prediction of treatment responses, and individualized risk assessment of retinal diseases, particularly in common conditions such as diabetic retinopathy and age-related macular degeneration. Despite these promising advances, the clinical translation of radiomics remains challenging. Current limitations include suboptimal model performance and generalizability,as well as insufficient clinical interpretability of radiomic feature and predictive models, which hampers their integration into existing imaging systems and routine clinical workflows. Based on a systematic analysis of relevant articles published over the past five years, this paper summarizes recent progress in the application of radiomics combined with machine learning for the diagnosis and prognostic assessment of retinal diseases, and to discuss the key challenges and future directions for its clinical implementation.