MLB-Net: A Multi-level Lesion-Aware and Boundary-Enhanced Network for Polyp Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate automatic polyp segmentation is vital for early colorectal cancer diagnosis.
• Existing methods struggle with variations in polyp size, shape, and number, especially small or multiple polyps.

Purpose of the Study:

• To introduce MLB-Net, a novel network designed to overcome limitations in cross-level feature interaction and boundary modeling for polyp segmentation.
• To enhance the accuracy and robustness of automatic polyp segmentation.

Main Methods:

• Proposed MLB-Net, a Multi-level Lesion-aware and Boundary-enhanced Network.
• Utilized a pyramid vision Transformer for multi-scale feature extraction.
• Introduced Multi-level Position and Boundary Fusion (MPDF), Selective Step Feature Aggregation (SSFA), and Multi-level Detail Injection (MDI) modules.

Main Results:

• MLB-Net demonstrated superior performance across five public datasets compared to state-of-the-art methods.
• Achieved high mDice scores: 92.6% on Kvasir-SEG and 94.5% on CVC-ClinicDB.
• Showcased effectiveness in segmenting diverse polyp appearances, including small-scale and multiple instances.

Conclusions:

• MLB-Net significantly improves polyp segmentation accuracy, especially for challenging cases.
• The model holds strong potential for clinical application in computer-aided colorectal cancer diagnosis.
• The proposed architecture effectively integrates multi-level features and enhances boundary details.