Deep learning for retina structural biomarker classification using OCT images

Abstract

This study presents an approach to identifying retinal structural biomarkers in ophthalmology, which is essential for accurate diagnosis and effective treatment of eye diseases. We develop a multi-modal, multi-task deep learning framework that integrates supervised and semi-supervised training methods. This model effectively processes a combination of 3D Optical Coherence Tomography (OCT) images and one-dimensional clinical data. A key advancement is introducing a custom post-processing method that significantly improves the precision of biomarker detection. Our model successfully identifies six distinct biomarkers in the retina and achieves a notable macro f1-score of 71.62%, representing a substantial 14.48% improvement over the baseline performance. This advancement underscores the potential of deep learning in enhancing diagnostic accuracy and treatment efficacy in ophthalmology.