InCoLoTransNet: An Involution-Convolution and Locality Attention-Aware Transformer for Precise Colorectal Polyp...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Gastrointestinal disease diagnosis is challenging due to complex anatomy.
• Current methods like colonoscopy generate large datasets, requiring time-consuming manual analysis.
• Automated systems are needed to assist clinicians in low-cost, effective disease identification.

Purpose of the Study:

• To introduce InCoLoTransNet, a novel framework for accurate polyp segmentation.
• To develop a computer-assisted system for efficient analysis of gastrointestinal endoscopic data.
• To improve the decision-making process for clinical professionals in diagnosing GI diseases.

Main Methods:

• Utilized a novel InCoLoTransNet framework with an encoder-decoder architecture.
• Employed a vision transformer in the encoder for global context and convolution-involution in the decoder for feature resampling.
• Integrated CBAM and locality self-attention modules to refine and capture relevant spatial and contextual information.

Main Results:

• InCoLoTransNet achieved optimal polyp segmentation performance across five public datasets.
• The framework attained the highest mean dice score (93%) on CVC-ColonDB and mean intersection over union (90%).
• Demonstrated strong generalization performance on unseen datasets, with high scores in mean dice coefficient and mean intersection over union.

Conclusions:

• InCoLoTransNet significantly outperforms 15 state-of-the-art polyp segmentation methods.
• The framework offers enhanced accuracy and generalization capabilities for computer-assisted GI disease detection.
• InCoLoTransNet represents a valuable tool for improving the efficiency and effectiveness of clinical decision-making in gastroenterology.