Mixture-attention Siamese transformer for video polyp segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp segmentation in colonoscopy is crucial for colorectal cancer prevention and treatment.
• Modeling long-range spatio-temporal relationships in videos presents a significant challenge for current methods.

Purpose of the Study:

• To develop a novel deep learning model for accurate polyp segmentation in colonoscopy videos.
• To address the challenge of modeling long-range spatio-temporal dependencies within colonoscopy data.

Main Methods:

• A Mixture-Attention Siamese Transformer (MAST) architecture was proposed.
• The model utilizes a Siamese transformer to encode paired video frames and a mixture-attention module to capture intra-frame and inter-frame correlations.
• Enhanced features are processed by parallel decoders for segmentation map prediction.

Main Results:

• MAST demonstrated superior performance on the SUN-SEG benchmark compared to existing state-of-the-art methods.
• The model effectively captures long-range spatio-temporal relationships for improved segmentation accuracy.

Conclusions:

• The proposed MAST model offers a significant advancement in polyp segmentation for colonoscopy videos.
• This approach holds promise for enhancing early detection and treatment of colorectal cancer.