BFE-Net: bilateral fusion enhanced network for gastrointestinal polyp segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of gastrointestinal polyps is crucial for diagnosis and treatment.
• Existing methods face challenges with small polyps or those resembling surrounding tissues.
• Pyramid Vision Transformer (PVT) captures global context but may miss details, while U-Net excels at semantic extraction.

Purpose of the Study:

• To develop an advanced deep learning model for improved gastrointestinal polyp segmentation.
• To address limitations of current segmentation techniques, particularly for challenging polyp cases.
• To enhance the accuracy and reliability of polyp detection in endoscopic imaging.

Main Methods:

• Proposed the bilateral fusion enhanced network (BFE-Net).
• Integrated U-Net and Pyramid Vision Transformer (PVT) features using a deep feature enhancement fusion module (FEF).
• Incorporated an attention decoder module (AD) for refined feature processing.

Main Results:

• BFE-Net demonstrated significant improvements in polyp segmentation accuracy.
• The model showed effectiveness across diverse datasets and imaging modalities.
• Validation confirmed the model's capability in handling difficult polyp segmentation scenarios.

Conclusions:

• The proposed BFE-Net effectively addresses key challenges in gastrointestinal polyp segmentation.
• The fusion of U-Net and PVT features with attention mechanisms enhances diagnostic capabilities.
• This advancement holds promise for improving gastrointestinal polyp diagnosis and treatment outcomes.