SAM-BoMu: Boundary-Aware Multi-Channel Fusion Network for Precise Polyp Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate colorectal polyp segmentation is crucial for colorectal cancer diagnosis and treatment.
• The Segment Anything Model (SAM) shows promise in medical imaging but struggles with polyp variability and boundary ambiguity.
• Existing methods face challenges due to domain differences between natural and medical images.

Purpose of the Study:

• To propose SAM-BoMu, a SAM-based boundary-aware multi-channel fusion network for precise colorectal polyp segmentation.
• To address segmentation challenges posed by polyp size/shape variability and boundary ambiguity.
• To reduce the training cost of the Vision Transformer (ViT) encoder using an Optimization-Aid Framework for Transformer (OAFT).

Main Methods:

• Developed SAM-BoMu, featuring a Multi-Channel Encoder (MCE) with ViT, CNN, and boundary extractors, and a Multi-Fusion Decoder (MFD).
• Integrated an Optimization-Aid Framework for Transformer (OAFT) to efficiently train the ViT encoder by freezing most parameters.
• Employed a Boundary-Constrained Hybrid Fusion (BCHF) module in the MFD to leverage extracted features for enhanced segmentation.

Main Results:

• SAM-BoMu achieved competitive, state-of-the-art performance across five public datasets.
• Attained a Dice coefficient of 0.954 on the CVC-ClinicDB dataset, the highest reported.
• Demonstrated strong generalization capability with a Dice coefficient of 0.943 on a private dataset (AHJU-DB).

Conclusions:

• SAM-BoMu effectively overcomes SAM's limitations in medical image segmentation, particularly for colorectal polyps.
• The proposed network demonstrates superior performance and generalization ability in polyp segmentation tasks.
• This approach holds significant potential for improving colorectal cancer diagnosis and treatment through accurate polyp identification.