CAFE-Net: Cross-Attention and Feature Exploration Network for polyp segmentation

Abstract

Colorectal polyp segmentation can help physicians screen colonoscopy images, which is essential for preventing colorectal cancer. The segmentation of polyps encounters multiple challenges, like small size, uneven brightness, blurred edges, and the potential confusion of folds with objects. Although existing methods have shown promising performance in addressing these challenges, there are three main shortcomings: (1) during the segmentation process, small polyp objects are lost, (2) the decoder stage faces challenges in restoring fine-grained details of the features, and (3) the limited capacity to aggregate multi-scale features. We propose a cross-attention and feature exploration network (CAFE-Net) for polyp segmentation to address these challenges. The work offers the following contributions: (1) a feature supplement and exploration module (FSEM) supplements in missing details and explores latent features, (2) a cross-attention decoder module (CADM) effectively preserves features from lower layers and restores fine-grained information, and (3) a multi-scale feature aggregation (MFA) module maximizes the utilization of previously learned features. We conducted extensive experiments and compared CAFE-Net with nine state-of-the-art (SOTA) methods. The CAFE-Net demonstrates the best segmentation accuracy across multiple polyp datasets as well as has an obvious advantage in segmenting small polyp objects.