A Novel Application of a Semi-Supervised Learning Approach to Identify Choroidal Tumors in Fundus Photographs

Machine learning and deep learning have become very popular in solving many challenges in a wide range of applications in many areas of study. Healthcare is one of the areas in which machine learning and deep learning algorithms and techniques are making long strides in solving many of the challenges that face this sector. Choroidal tumors are a rare occurrence but affect patients in varying degrees depending on the type of tumor. Correctly discerning between malignant and nonmalignant choroidal lesions is critical in proper patient care. In this paper, we attempt to automate this process by applying a novel semi-supervised learning approach to a fundus photograph dataset. Multicon [1] is a semi-supervised learning method that utilizes stochastic data augmentation, pseudo labeling, and multicontrastive loss function that combines two loss functions: consistency regularization loss, and unsupervised multicontrastive loss. By applying this novel technique, we were able to achieve promising results.