Reliable classification of polyps based on artificial intelligence: a development and validation study

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology and machine learning

Background:

• Pathologist shortages create diagnostic bottlenecks.
• An increasing volume of colorectal biopsies requires efficient diagnostic tools.
• Artificial intelligence offers a solution for prescreening colorectal biopsies.

Purpose of the Study:

• To develop and validate an AI-based prescreening tool, POLARIS, for colorectal biopsies.
• To assist pathologists in managing the growing caseload of colorectal biopsies.
• To improve the accuracy and efficiency of colorectal polyp diagnosis.

Main Methods:

• Developed POLARIS using a foundation model (H-optimus-0) and multiple instance learning on 15,079 whole-slide images (WSIs).
• Trained and validated the model on datasets from the UK bowel cancer screening program and Cheltenham General Hospital.
• Classified WSIs into risk categories and validated performance using geographically external datasets and expert pathologist review.

Main Results:

• POLARIS achieved 98.94% accuracy in identifying high-grade dysplasia (HGD) and adenocarcinoma, and 83.04% accuracy in classifying normal/low-grade dysplasia (LGD) cases.
• The model demonstrated a balanced accuracy of 86.65% in external validation and an AUROC of 0.9449 for distinguishing LGD from polyps needing review.
• Expert pathologists agreed with POLARIS in 92.5% of challenging cases, and AI-generated heatmaps correlated with pathologist-identified high-risk areas.

Conclusions:

• POLARIS effectively prescreens colorectal biopsies, identifying high-risk lesions with high accuracy.
• The AI tool can significantly reduce the number of slides requiring pathologist review, enhancing workflow efficiency.
• POLARIS shows potential to improve diagnostic turnaround times and support pathologists in managing large caseloads.