A Probabilistic Model to Support Radiologists' Classification Decisions in Mammography Practice

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Mammography is crucial for breast cancer screening.
• Radiologist classification decisions can exhibit variability.
• The Breast Imaging Reporting and Data System (BI-RADS) provides standardized classification guidelines.

Purpose of the Study:

• To develop a probabilistic model to aid radiologist classification decisions in mammography.
• To understand and quantify variability in radiologist judgment thresholds.
• To assess the potential of a model to improve accuracy and consistency in mammographic interpretation.

Main Methods:

• Trained a probabilistic Bayesian network (BN) on 112,433 mammographic cases.
• Modeled radiologist feature observations and BI-RADS classifications.
• Learned individual radiologist probabilistic thresholds within the BN model.
• Compared BN model performance against radiologists using observed vs. standard BI-RADS thresholds.

Main Results:

• Significant variability in radiologist observed thresholds was identified.
• Using observed thresholds, the BN model reduced false positives by 28.9% with a 0.01% increase in false negatives.
• Using the standard BI-RADS threshold, the BN model reduced false positives by 47.3% but increased false negatives by 26.7%.

Conclusions:

• A probabilistic model can significantly reduce false positives in screening mammography with minimal increase in false negatives.
• Learning radiologist thresholds offers insights into clinical practice conservativeness and performance variability.
• The developed model can support radiologists, enhancing performance and consistency in mammography interpretation.