Multi-Scale Cross-Attention Multiple Instance Learning Network for Automated Classification of Colorectal Polyps

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Gastrointestinal Pathology

Background:

• Gastrointestinal (GI) pathology services handle a high volume of colon polyp specimens, representing 40% of cases.
• Updated colorectal cancer screening guidelines lower the screening age to 45, potentially increasing endoscopy procedures and specimen volume.
• Developing an artificial intelligence (AI) classifier is crucial for efficient triage of colorectal polyp specimens.

Purpose of the Study:

• To develop and evaluate an AI classifier for triaging colorectal polyp specimens.
• To perform binary (neoplastic vs. nonneoplastic) and 12-way (final diagnosis) classifications.
• To assess the AI classifier's performance on archived, routine clinical, and external datasets.

Main Methods:

• Retrospective analysis of 1191 colon polyp slides (2021-2023).
• Training a multi-scale cross attention multiple instance learning (MsCAMIL) network using weakly-supervised transformer-based models.
• Utilizing Leica Aperio scanners for slide digitization and evaluating performance metrics including F1-Score and accuracy.

Main Results:

• The AI classifier achieved >95% accuracy and F1-score in binary classification for archived and routine clinical cases.
• 12-way classification yielded lower F1 scores (74% archived, 57% routine).
• Performance on an external dataset decreased to 86% accuracy, with variations noted across institutions.

Conclusions:

• MsCAMIL networks can serve as an efficient triage system in daily clinical workflow for colon polypectomy specimens.
• The AI classifier demonstrates high specificity and accuracy (>95%) in binary classification for both archived and routine cases.
• Multi-institutional collaborations are essential for further validation and broader clinical implementation.