Automatic deep learning-based colorectal adenoma detection system and its similarities with pathologists.

Area of Science:

• Digital pathology and computational diagnostics.
• Artificial intelligence in medical imaging analysis.
• Gastrointestinal pathology and cancer screening.

Background:

• The digital transformation of slide evaluation enables computer-aided diagnosis.
• Understanding deep learning (DL) model reasoning is crucial for clinical integration.
• Colorectal adenoma diagnosis offers a suitable model for AI evaluation due to its defined criteria.

Purpose of the Study:

• To compare the diagnostic performance and reasoning of a DL model with human pathologists in colorectal adenoma detection.
• To assess the generalizability of the DL model across different data sources.

Main Methods:

• A DL model (DeepLab v2 with ResNet-34) was trained on 177 annotated slides.
• Model performance was evaluated on 194 test slides and compared against five pathologists.
• Generalization was tested using 168 additional slides from two other hospitals.

Main Results:

• The DL model achieved an AUC of 0.92 and >90% slide-level accuracy on external datasets.
• Model performance was comparable to experienced pathologists and superior to the average pathologist.
• Analysis of feature maps and misdiagnosed cases revealed similar diagnostic reasoning processes.

Conclusions:

• DL models demonstrate comparable performance and reasoning to pathologists in colorectal adenoma diagnosis.
• The DL model exhibits robustness, maintaining high accuracy despite variations in staining and slide origin.
• These findings support the potential of DL as a valuable tool in computer-aided diagnosis for pathology.