MNet-SAt: A Multiscale Network with Spatial-enhanced Attention for segmentation of polyps in colonoscopy

Abstract

Automated polyp segmentation in colonoscopy is crucial for early diagnosis and treatment planning of colorectal cancer. Early detection of polyps significantly improves survival rates. However, traditional methods rely on manual verification by endoscopic physicians, increasing the risk of missed polyps due to their variability. To address this, we propose a novel Multiscale Network with Spatial-enhanced Attention (MNet-SAt) for automated polyp segmentation. MNet-SAt preserves edge information to improve polyp boundary detection and efficiently utilizes multi-scale features across spatial and channel dimensions, focusing on key regions. Additionally, a spatially enhanced attention mechanism captures spatial relationships and global context, further improving segmentation accuracy. We comprehensively evaluate the efficacy of MNet-SAt on the Kvasir-SEG and CVC-ClinicDB datasets. Qualitative and quantitative experiments demonstrate the superiority of MNet-SAt over ten state-of-the-art existing baselines and its stronger generalization performance. On Kvasir-SEG, it achieves an outstanding performance of 96.61% DSC, 96.92% IoU, 97.36% Precision, 97.83% Accuracy, and 97.12% Recall. On CVC-ClinicDB, it achieves 98.60% DSC, 95.89% IoU, 96.89% Precision, 97.32% Accuracy, and 96.13% Recall. In summary, MNet-SAt offers remarkable performance in automating polyp segmentation. The superior performance of MNet-SAt shows great potential for improving the diagnostic process in early polyp identification and more effective treatment planning, thereby contributing to lowering colorectal cancer mortality rates. © 2024 Elsevier Ltd