Automated classification of polypoidal choroidal vasculopathy and wet age-related macular degeneration by spectral domain optical coherence tomography using self-supervised learning

Abstract

This work aimed to evaluate the performance of deep learning algorithms for an automated differentiation between polypoidal choroidal vasculopathy (PCV) and wet age-related macular degeneration (wet-AMD) by optical coherence tomography (OCT) images of macula. The automated classification model was developed with the self-supervised learning technique and was analyzed for the model performance. It learned visual representations from unlabeled data prior to fitting with labeled data. This technique was shown to provide desirable performance based only on a small proportion of labeled data. The final ensemble model trained with this technique with selected parameters yielded promising performance. Particularly, the proposed model provided 0.71 AUC with 0.67 sensitivity and 0.8 specificity on test set which were better than the traditional supervised learning models used as baseline approached in this study.