Polyp segmentation for colonoscopy images via Hierarchical Interworking Decoding

Abstract

Efficient and accurate identification, localization, and segmentation of polyp tissues are critical steps in colonoscopy, essential for the prevention and early intervention of colorectal cancer. Current CNN-based methods are limited in modeling long-range dependencies while transformer-based methods cannot capture sufficient contextual dependencies. Hybrid networks are prone to overfitting the convolutional features, leading to the dispersion of attention in the Transformer. Addressing the existing issues, we propose an approach for polyp segmentation with Hierarchical Interworking Decoder (HID) that fully utilizes hierarchical features to establish multi-scale discriminative criteria. HID leverages Interworking Attention Module (IAM) to refine single-level features, where the globally shared attention mechanism in IAM concurrently integrates affinity information from all different hierarchical features, facilitating global information exchange. Adjacent Aggregation Module (AAM) to refine and integrate adjacent-level features. Through the refinement of single-level features and the integration of different-level features, HID simultaneously captures global information and local contextual information. Extensive experiments demonstrate that HID exhibits outstanding generalization performance and achieves state-of-the-art accuracy on multiple polyp segmentation benchmarks.