HiT: holistic-integration transformer network for efficient segmentation in endoscopic medical imaging

Abstract

In recent years, accurate and efficient polyp segmentation has become essential for the early screening and diagnosis of gastrointestinal cancers, providing critical support for improving clinical outcomes. With the rapid advancement of hybrid Transformer-based methods, endoscopic polyp segmentation technology has achieved remarkable progress. Nevertheless, this task remains highly challenging due to factors such as irregular polyp morphology, substantial size variation, complex background textures, and ambiguous boundaries between lesions and surrounding tissues, all of which hinder precise identification and detailed analysis. To address these challenges, this paper proposes the HiT framework, a novel architecture specifically designed for endoscopic polyp segmentation to enhance both accuracy and robustness. The framework is built upon two key components: a hierarchical context fusion module and a shallow semantic fusion module. Through their coordinated interaction, the proposed design enables effective multi-scale contextual modeling and precise integration of spatial feature representations, thereby improving feature extraction efficiency and cross-level feature fusion capability. Extensive qualitative and quantitative experiments conducted on three public benchmark datasets demonstrate the effectiveness of the HiT framework. The proposed method achieves superior performance across multiple evaluation metrics, including HD95, Dice coefficient, mean intersection-overunion, precision, and recall. These results highlight the importance of accurate boundary delineation and comprehensive contextual representation in advancing endoscopic image segmentation technology, ultimately contributing to more reliable computer-aided diagnosis and clinical decision-making.