G-SET-DCL: a guided sequential episodic training with dual contrastive learning approach for colon segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate colon segmentation is crucial for effective CT colonography.
• Limited annotated data poses a significant challenge in medical image analysis.
• Deep learning offers potential for improving segmentation accuracy.

Purpose of the Study:

• To introduce a novel deep learning approach for enhanced colon segmentation.
• To improve the accuracy of CT colonography by addressing data limitations.
• To boost the clinical effectiveness of the CT colonography pipeline.

Main Methods:

• Integration of 3D contextual information using guided sequential episodic training.
• Rectum detection via a Markov Random Field-based algorithm.
• Supervised sequential episodic training and contrastive learning for feature enhancement.

Main Results:

• Achieved a 97.3% Dice coefficient and 98.28% polyp information preservation accuracy on 98 scans.
• Demonstrated robustness and reliability with statistical analysis (95% confidence intervals).
• Maintained high accuracy (97.15% Dice coefficient) even with limited annotated data (10 scans).

Conclusions:

• The sequential segmentation approach shows promising results for colon segmentation.
• The method exhibits strong generalization capabilities, particularly with limited annotated datasets.
• This approach addresses a common challenge in medical imaging, enhancing diagnostic potential.