Multi-Modality Semi-Supervised Learning for Ophthalmic Biomarkers Detection

Ophthalmic Biomarkers, as an objective and quantifiable approach to identifying the ophthalmological disease process, are proven to be useful not only in assisting healthcare professionals in disease diagnosis but also in the identification of phenomena and risk factors in the early stages, which greatly contribute to disease prevention and better treatment of patients. In this study, a deep learning method is introduced to achieve simultaneous automatic recognition of six prevalent ophthalmic biomarkers in the OLIVES dataset. To enhance identification accuracy, semi-supervised learning techniques are adopted in this research and different data modalities are jointly optimized using a guided loss function. The experimental results reveal that the method reaches an F1 score of 0.70 on a test set with 3,872 images.