A lighter hybrid feature fusion framework for polyp segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Colonoscopy is vital for colorectal cancer screening, but polyp identification and segmentation are challenging due to variations in polyp appearance.
• Deep learning offers efficient polyp segmentation, yet current Transformer-based methods struggle with local details and computational cost.
• Accurate polyp segmentation is crucial for early detection and effective treatment of colorectal cancer.

Purpose of the Study:

• To develop a novel CNN-Transformer hybrid model (CTHP) for improved polyp segmentation in colonoscopy images.
• To address the limitations of existing Transformer models, including overlooking local details and high computational burden.
• To enhance the accuracy and efficiency of polyp segmentation for early colorectal cancer detection.

Main Methods:

• Proposed a CNN-Transformer hybrid model (CTHP) integrating CNN's local feature extraction with Transformer's global semantic understanding.
• Optimized Transformer's self-attention mechanism by computing it along width and height directions for improved computational efficiency.
• Introduced an information propagation module and positional bias coefficients to mitigate information dispersal during feature fusion.

Main Results:

• The CTHP model achieved state-of-the-art performance on multiple benchmark datasets for polyp segmentation.
• Demonstrated superior accuracy and efficiency compared to existing methods in colonoscopy image analysis.
• Exhibited excellent cross-domain generalization performance, indicating robustness across different datasets.

Conclusions:

• The proposed CTHP model effectively combines CNN and Transformer strengths for accurate and efficient polyp segmentation.
• CTHP offers a promising solution for overcoming the limitations of current deep learning approaches in medical image analysis.
• The model's strong performance and generalization capabilities support its potential application in clinical settings for colorectal cancer screening.