Dual multi scale networks for medical image segmentation using contrastive learning

Abstract

DMSNet, a novel model for medical image segmentation is proposed in this research work. DMSNet employs a dual multi-scale architecture, combining the computational efficiency of EfficientNet B5 with the contextual understanding of the Pyramid Vision Transformer (PVT). Integration of a multi-scale module in both encoders enhances the model's capacity to capture intricate details across various resolutions, enabling precise delineation of complex foreground boundaries. Notably, DMSNet incorporates contrastive learning with a novel pixel-wise contrastive loss function during training, contributing to heightened segmentation accuracy and improved generalization capabilities. The model's performance is demonstrated through experimental evaluation on the four diverse datasets including Brain tumor segmentation (BraTS 2020), Diabetic Foot ulcer segmentation (DFU), Polyps (KVASIR-SEG) and Breast cancer segmentation (BCSS). We have employed recently introduced metrics to evaluate and compare our model with other state-of-the-art architectures. By advancing segmentation accuracy through innovative architectural design, multi-scale modules, and contrastive learning techniques, DMSNet represents a significant stride in the field, with potential implications for improved patient care and outcomes.