Deep learning system for true- and pseudo-invasion in colorectal polyps.

Area of Science:

• Pathology
• Artificial Intelligence
• Computational Biology

Background:

• Over 15 million colonoscopies are performed annually in North America.
• Accurate differentiation between true and pseudo-invasion in colon polyps is crucial for appropriate patient treatment.
• Current diagnostic methods lack specialized tools and well-annotated datasets, presenting a challenge for pathologists.

Purpose of the Study:

• To develop and evaluate a novel AI system for classifying tissue types and differentiating true- versus pseudo-invasion in colon polyps.
• To create an online annotation tool to facilitate the training of deep neural networks using whole-slide images (WSIs).

Main Methods:

• Acquired 150 whole-slide images (WSIs) from colon biopsies.
• Developed three deep neural networks (DNNs) trained on WSIs at different magnifications.
• Utilized an online annotation platform for pathologist input to train the DNNs.

Main Results:

• The AI system achieved 95.3% accuracy in classifying tissue types.
• The system demonstrated 83.9% accuracy in differentiating true- versus pseudo-invasion.
• The system's diagnostic efficiency was found to be comparable to that of an expert pathologist.

Conclusions:

• The developed AI system offers a promising tool for assisting pathologists in diagnosing colon polyps.
• The system can function as a confirmatory or screening tool, potentially improving patient care and reducing healthcare costs.
• The availability of this system (http://ai4path.ca) marks a significant advancement in computational pathology for gastrointestinal diagnostics.