RSAFormer: A method of polyp segmentation with region self-attention transformer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonoscopy is crucial for early colon cancer detection.
• Accurate polyp segmentation in colonoscopy images is challenging due to indistinct boundaries.
• Existing models struggle with precise segmentation of polyps.

Purpose of the Study:

• To develop an advanced network, RSAFormer, for enhanced colon polyp segmentation.
• To improve the accuracy of identifying polyp boundaries in colonoscopic images.
• To address limitations in current state-of-the-art segmentation models.

Main Methods:

• Proposed RSAFormer, a network utilizing a transformer encoder for robust feature capture.
• Implemented a unique dual-decoder structure for flexible and detailed feature extraction.
• Introduced a region self-attention enhancement module (RSA) to refine uncertain areas and boundary information.

Main Results:

• RSAFormer demonstrated superior performance on five polyp datasets.
• Achieved a mean Dice score of 92.2% on the Kvasir dataset.
• Achieved a mean Dice score of 83.5% on the ETIS dataset, outperforming existing models.

Conclusions:

• RSAFormer effectively enhances feature representation and boundary delineation for colon polyp segmentation.
• The dual-decoder and region self-attention mechanisms contribute to improved segmentation accuracy.
• The proposed method shows significant potential for clinical application in early colon cancer screening.