iMIL4PATH: A Semi-Supervised Interpretable Approach for Colorectal Whole-Slide Images

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Colorectal cancer (CRC) diagnosis relies on manual pathology review of biopsies.
• Increasing CRC incidence and workload necessitate efficient diagnostic tools.
• Automated analysis of whole-slide images (WSI) can support pathologists.

Purpose of the Study:

• To develop an interpretable semi-supervised AI model for detecting colorectal cancer lesions in biopsies.
• To improve diagnostic accuracy and efficiency in digital pathology workflows.
• To assist pathologists in case triage and review of WSI.

Main Methods:

• Implemented a semi-supervised learning approach using multiple-instance learning and feature aggregation.
• Trained and evaluated the model on an extended CRC dataset (CRC+) comprising 4433 WSI.
• Assessed model performance using classification accuracy, sensitivity, specificity, and quadratic weighted kappa.

Main Results:

• Achieved 90.19% classification accuracy and 98.8% sensitivity for CRC lesion detection.
• Obtained 85.7% specificity and a quadratic weighted kappa of 0.888.
• Demonstrated generalization capabilities on two external public datasets.

Conclusions:

• The proposed interpretable AI model effectively detects colorectal cancer lesions in biopsies with high sensitivity.
• This approach can significantly aid pathologists in managing increased workloads and improving diagnostic accuracy.
• The model shows promise for integration into clinical digital pathology workflows.