Broad-spectrum eye disease classification using a deep learning-based tailored software lens

Abstract

The early and accurate classification of eye diseases is essential for preventing irreversible visual impairment. This task can be performed by deep learning approaches that automatically classify retinal fundus images according to potential illnesses. Despite notable advances in this field, the robust and methodologically rigorous classification of a broad range of eye diseases remains unsolved. This study addresses this issue by proposing a novel deep learning architecture that leverages specific features of retinal fundus images (e.g., image noise and importance of fine structures) using a tailored software lens to robustly diagnose a broad spectrum of illnesses at a high performance level. To validate this approach, the currently broadest peer-reviewed dataset of 16,242 images, comprising nine diseases and healthy samples, is chosen. Our novel architecture achieves a 5-fold cross-validated average balanced accuracy of 82.52 %, outperforming the baseline model (79.40 %) and setting a new benchmark. Our results demonstrate for the first time that high performance can be achieved for diagnosing a broad range of eye diseases based on retinal fundus images by leveraging their specific features. This approach has implications for clinical deployment, particularly in routine care settings, by enabling faster and more reliable screenings.