Classification of Polyps in Endoscopic Images Using Self-Supervised Structured Learning

This study uses a two-stage learning computer-aided diagnosis (CAD) scheme that has a convolutional neural network(CNN) with self-supervised learning(SSL) to classify polyps as either a hyperplastic polyp (HP) or a Tubular Adenoma (TA). The proposed model uses look-into-object (LIO) and contrastive learning in SimCLR to focus on the holistic polyp region and allows greater model performance. However, the LIO scheme relies on pretraining a model to provide basic representations so this model is modified using a warm-up scheme to improve the loss function. There are insufficient medical images to train efficient representation for polyp classification so another approach uses natural images, instead of polyp images, for the pretext task. The experimental results show that the proposed scheme which uses polyp object structure information and self-supervised learning produces a robust model that allows better classification as either HP or TA in the prediction head by transferring a backbone. The backbone model uses ResNet-18 effectively to concentrate on the holistic polyp using limited labeled polyp images. The proposed scheme outperforms an existing method with a 4% increase in accuracy and a 3% improvement in F1-score.