University of Washington’s Explainable AI Model Outperforms Benchmarks in Breast MRI Tumor Localization
Its pixel-level anomaly maps have enhanced radiologist interpretability in diverse patient cohorts
Overview
- The model was trained on nearly 10,000 contrast-enhanced breast MRI exams from 2005 to 2022 at the University of Washington to reflect real-world cancer prevalence
- In both internal and external tests the AI accurately highlighted biopsy-proven tumors and surpassed traditional binary classifiers in high- and low-prevalence screening scenarios
- Spatially resolved heatmaps provide clear, pixel-level explanations of suspect regions, matching radiologists’ annotations and boosting clinical trust
- Unlike earlier tools, this anomaly detection approach uses realistic distributions of normal and malignant cases and prioritizes explainability for potential clinical adoption
- Researchers plan larger prospective studies to evaluate integration into radiology workflows for triaging normal scans and improving reading efficiency