MBRSNet: Boundary-Aware Multi-Task Learning with Signed Distance Field Regression for Polyp Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate polyp segmentation in colonoscopies is crucial but challenging due to image quality and data variability.
• Existing methods struggle with integrating boundary and region information, limiting segmentation accuracy and generalization.

Purpose of the Study:

• To develop a boundary-aware multi-task learning framework (MBRSNet) for improved colonoscopic polyp segmentation.
• To explicitly model the synergy between segmentation and boundary prediction tasks.

Main Methods:

• Proposed MBRSNet framework utilizing multi-task learning with segmentation and signed distance field (SDF) regression.
• Introduced a cross-gated multi-task bottleneck for selective feature interaction between tasks.
• Implemented a hierarchical cross-task guidance strategy in the decoder for feature refinement.

Main Results:

• MBRSNet achieved competitive or superior performance on five benchmark datasets compared to state-of-the-art methods.
• Demonstrated enhanced boundary delineation accuracy, especially under challenging conditions.
• Showcased strong robustness to domain shifts and improved cross-dataset generalization.

Conclusions:

• The proposed MBRSNet framework effectively integrates boundary and regional information for robust polyp segmentation.
• Structured task interaction via multi-task learning significantly improves segmentation accuracy and generalization in colonoscopic images.
• MBRSNet offers a promising solution for boundary-aware medical image segmentation challenges.