AI Should Read Mammograms Only When Confident: A Hybrid Breast Cancer Screening Reading Strategy.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Breast Cancer Screening

Background:

• Quantifying uncertainty in AI-based mammogram interpretations is crucial for integrating AI into screening programs.
• Current AI models require evaluation for their impact on radiologist workload and diagnostic performance.

Purpose of the Study:

• To assess the potential reduction in radiologist workload during mammographic screening.
• To evaluate the performance of an AI model incorporating uncertainty quantification in mammography interpretation.
• To determine if AI integration maintains or improves cancer detection and recall rates.

Main Methods:

• Developed an AI model outputting probability of malignancy (PoM) and uncertainty.
• Implemented a hybrid reading approach: AI handles confident predictions, radiologists review uncertain cases.
• Retrospectively optimized and tested the approach on a large dataset (41,469 examinations) using various uncertainty metrics.

Main Results:

• The best uncertainty metric (entropy of mean PoM) allowed AI to read 61.9% of examinations.
• Hybrid reading achieved recall (23.6‰) and cancer detection rates (6.6‰) similar to standard double reading.
• AI model's AUC was significantly lower for uncertain (0.87) versus certain (0.96) predictions (P=.02).

Conclusions:

• AI mammography interpretation with uncertainty quantification can substantially decrease radiologist workload.
• This approach maintains comparable cancer detection and recall rates to traditional double reading.
• Uncertainty quantification enables effective AI integration, optimizing screening efficiency.