CMNet: deep learning model for colon polyp segmentation based on dual-branch structure.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colon cancer is a leading gastrointestinal cancer, with colon polyps as a primary precursor.
• Early detection and removal of colon polyps are crucial for preventing colon cancer incidence.
• Artificial intelligence (AI) and deep learning are increasingly utilized in medical diagnosis to aid clinicians.

Purpose of the Study:

• To develop and validate a deep learning model for accurate colon polyp segmentation.
• To improve the early diagnosis of colon polyps, thereby reducing the risk of colon cancer.
• To leverage advanced AI for enhanced medical diagnosis and treatment planning.

Main Methods:

• A dual-branch deep learning model combining Convolutional Neural Networks (CNNs) and transformers.
• Utilized deeply separable convolution based on ResNet and incorporated a stripe pooling module.
• Introduced an aggregated attention module (AAM) for high-dimensional semantic information fusion.
• Employed deep supervision and multi-scale training to enhance model performance and generalization.

Main Results:

• The proposed dual-branch structure significantly outperformed single-branch models.
• The model incorporating the aggregated attention module (AAM) showed substantial performance improvements.
• Achieved leading results on the Kvasir-SEG dataset, with 1.1% higher mIoU and 1.5% higher mDice compared to state-of-the-art models.

Conclusions:

• Validated a novel deep learning model for colon polyp segmentation using a dual-branch network.
• Demonstrated the effective complementarity of CNNs and transformers in this application.
• Confirmed the feasibility of fusing different structures for high-dimensional semantics, retaining crucial information.