Multi-View Attention Network With Iterative Feature Refinement and Boundary Awareness for Endoscopic Image Segmentation

Endoscopic image segmentation plays a critical role in diagnosing early gastrointestinal tumors, which is essential for preventing colorectal and gastric cancers. However, achieving accurate segmentation is challenging due to issues such as boundary blurring, low contrast, and small lesion sizes. This paper proposes a novel Multi-view Attention Network (MVANet) to address these challenges by incorporating several specialized blocks. The first, a Composite Block (CMB) enhances feature representations across multiple dimensions, improving segmentation accuracy. To bridge the semantic gap between different feature layers, we propose an Attention-based Cross-layer Feature Fusion (ACFF) block, which incorporates a Triplet Efficient Transformer Attention (TETA) mechanism to capture long-range dependencies across multiple views. Additionally, to enhance the model’s boundary-awareness, the paper presents a Prior Knowledge-guided Boundary Awareness (PKBA) block, which aids in capturing irregular boundaries, and a Boundary-driven Scale Awareness and Semantic Complementarity (BSASC) block, designed to improve boundary localization during the decoder stage. Moreover, a Multi-scale Feature Integration (MFI) block is proposed to integrate multi-scale features during the decoder stage to capture potentially useful features. An iterative feature refinement module, composed of four Image-Guided Feature Refinement (IFR) blocks, is used to further refine local information. Extensive experiments conducted on six endoscopic image datasets demonstrate that MVANet achieves state-of-the-art performance, with mDice values of 93.1% and 94.9%, and mIoU values of 88.4% and 90.7% on the Kvasir-SEG and CVC-ClinicDB datasets, respectively. MVANet demonstrates strong potential for enhancing segmentation performance in early tumor detection.