Machine learning classification of diagnostic accuracy in pathologists interpreting breast biopsies

Area of Science:

• Digital pathology
• Machine learning in medicine
• Medical image analysis

Background:

• Pathologist diagnostic accuracy is crucial for patient outcomes.
• Digital pathology enables detailed analysis of whole slide images.
• Understanding pathologist viewing behavior can reveal cognitive processes.

Purpose of the Study:

• To assess the feasibility of using machine learning to predict diagnostic accuracy from pathologist viewing behavior.
• To identify key viewing behavior features that correlate with diagnostic accuracy.
• To develop predictive models for diagnostic accuracy in digital breast biopsy interpretation.

Main Methods:

• Collected viewing behavior data (zooming, panning) from 140 pathologists evaluating digital breast biopsy images.
• Extracted 30 features from viewing behavior and employed 4 machine learning algorithms.
• Utilized Random Forest classifier for predicting diagnostic accuracy.

Main Results:

• The Random Forest model achieved 0.81 test accuracy and 0.86 AUC.
• Attention distribution and focus on critical regions were significant predictors.
• Incorporating case-level and pathologist-level data further enhanced classifier performance.

Conclusions:

• Pathologist viewing behavior can predict diagnostic accuracy in digital pathology.
• Developed models offer potential for automated feedback and decision support systems.
• Findings have implications for training, clinical practice, and cognitive research.