MBFFNet: Multi-Branch Feature Fusion Network for Colonoscopy.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Colonoscopy is a primary tool for detecting rectal polyps and cancer.
• Deep learning-based semantic segmentation shows promise for medical lesion detection.
• Current methods struggle with high accuracy and real-time performance in polyp detection.

Purpose of the Study:

• To develop an accurate and real-time semantic segmentation method for colonoscopy polyp detection.
• To address the limitations of existing deep learning models in terms of accuracy and computational efficiency.

Main Methods:

• Proposed a multi-branch feature fusion network (MBFFNet) based on the UNet architecture.
• Implemented stepwise sampling with channel multiplication to reduce computational load.
• Utilized multi-layer feature extraction and varied resizing techniques to enhance information integration.

Main Results:

• MBFFNet demonstrated superior accuracy, reduced model size, and fewer computational operations (flops) compared to baseline models like UNet and UNet++.
• Achieved mIOU of 0.8952, F-score of 0.9450, and Dice loss with CE of 0.1602.
• Reduced flop count by 73.2% compared to UNet, with competitive performance against state-of-the-art PraNet.

Conclusions:

• MBFFNet offers an effective solution for accurate and real-time polyp segmentation in colonoscopy images.
• The model shows potential for broader applications in medical image segmentation tasks.
• MBFFNet balances high performance with significantly reduced computational requirements.