An interpretable machine learning system for colorectal cancer diagnosis from pathology slides

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Pathology practice is undergoing significant transformation.
• Accurate colorectal cancer diagnosis from whole-slide images (WSI) is crucial.
• Scalable AI solutions are needed to meet diagnostic demands.

Purpose of the Study:

• To develop a scalable artificial intelligence (AI) system for diagnosing colorectal cancer using whole-slide images (WSI).
• To improve diagnostic efficiency and accuracy in computational pathology.

Main Methods:

• A deep learning (DL) system trained with weak labels and a mixed-supervision scheme.
• An intelligent sampling strategy reducing training data by a factor of six without performance loss.
• Leveraging a small subset of fully annotated samples and incorporating explainable AI features.

Main Results:

• Achieved 93.44% accuracy and 0.996 sensitivity on an internal dataset of ~10,500 WSIs.
• Demonstrated robustness on external datasets (TCGA, PAIP), with 84.91% accuracy on TCGA.
• The mixed-supervision scheme enabled an effective intelligent sampling strategy.

Conclusions:

• The developed AI system offers a scalable and accurate solution for colorectal cancer diagnosis from WSI.
• The intelligent sampling strategy enhances training efficiency without compromising diagnostic performance.
• The AI system shows promise for integration into clinical pathology workflows.