Enhancing colorectal polyp segmentation with TCFMA-Net: A transformer-based cross feature and multi-attention network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer (CRC) diagnosis relies on accurate polyp segmentation in colonoscopy images.
• Existing segmentation methods face challenges with varying polyp sizes and precise boundary delineation.
• Early detection of polyps is crucial for effective CRC management and improved patient outcomes.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced polyp segmentation in colonoscopy.
• To address limitations in current methods regarding polyp size variability and boundary accuracy.
• To provide a robust tool for improving the accuracy and reliability of colorectal cancer diagnosis.

Main Methods:

• Proposed the Transformer-based cross feature multi-attention network (TCFMA-Net).
• Utilized swin transformer-based encoders for feature extraction.
• Incorporated a cross-feature enhancer network with multi-attention modules for comprehensive feature fusion.

Main Results:

• TCFMA-Net achieved superior polyp segmentation performance across six benchmark datasets.
• Demonstrated high Dice scores, e.g., 92.74 ± 0.10 on CVC-ClinicDB and 91.46 ± 0.14 on Kvasir-SEG.
• Showcased robust generalizability with strong performance on diverse datasets like CVC-300 and Polypgen.

Conclusions:

• TCFMA-Net significantly advances polyp segmentation technology for colonoscopy.
• The model effectively handles variations in polyp size and boundary complexities.
• Presents a reliable and accurate tool for early colorectal cancer detection and diagnosis.