BMSNet: a boundary-guided multi-scale polyp segmentation network

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp segmentation is crucial for computer-aided diagnosis and early colorectal cancer screening.
• Challenges include variations in polyp size, shape, texture, blurred boundaries, and low contrast.

Purpose of the Study:

• To propose BMSNet, a boundary-guided multi-scale network for accurate polyp segmentation.
• To address the limitations of existing methods in handling polyp variations and boundary ambiguity.

Main Methods:

• Utilized an SMT-T backbone for efficient multi-level feature extraction.
• Introduced a Boundary Prediction Module with bidirectional feature interaction and differentiable Canny-based supervision.
• Designed a Boundary-Guided Feature Enhancement Module incorporating frequency-domain enhancement, multi-scale feature interaction, and boundary-guided modulation.
• Employed a Feature Fusion Unit for progressive integration of multi-scale features.

Main Results:

• BMSNet achieved high Dice scores on benchmark datasets: 0.943 (ClinicDB), 0.920 (Kvasir), 0.828 (ColonDB), 0.825 (ETIS), and 0.913 (CVC-300).
• Demonstrated effectiveness across diverse datasets, indicating robust performance.
• Showcased a favorable balance between segmentation accuracy and computational efficiency.

Conclusions:

• BMSNet offers a promising solution for accurate polyp segmentation in colonoscopy images.
• The boundary-guided approach effectively enhances feature representation and segmentation accuracy.
• The method holds potential for improving computer-aided diagnosis and colorectal cancer screening.