Cancer Detection in Breast MRI Screening via Explainable AI Anomaly Detection

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial intelligence (AI) models offer potential for enhancing breast MRI screening accuracy and efficiency.
• Existing AI models lack rigorous evaluation in low-prevalence populations and interpretability, hindering clinical adoption.
• There is a need for explainable AI in breast MRI for reliable clinical use.

Purpose of the Study:

• To develop an explainable AI model for detecting cancer in breast MRI.
• To ensure the model's effectiveness in both high and low cancer prevalence scenarios.
• To improve the interpretability and clinical adoption of AI in breast MRI screening.

Main Methods:

• Developed an explainable fully convolutional data description (FCDD) anomaly detection model using 9567 breast MRI examinations.
• Evaluated model performance using cross-validation, an internal test set (171 exams), and an external multicenter dataset (221 exams).
• Assessed model explainability via pixelwise comparisons with malignancy annotations, using the Wilcoxon signed rank test for statistical significance.

Main Results:

• The FCDD model outperformed the benchmark binary cross-entropy (BCE) model in cross-validation for balanced (AUC=0.84 vs 0.81) and imbalanced (AUC=0.72 vs 0.69) tasks.
• In the imbalanced setting, FCDD achieved higher specificity (13% vs 9%) at 97% sensitivity.
• FCDD demonstrated superior spatial agreement with malignancy annotations (pixelwise AUC=0.92 vs 0.81) and strong performance on external testing (AUC=0.86 vs 0.79).

Conclusions:

• The developed explainable AI model accurately depicts tumor location in breast MRI.
• The FCDD model shows superior performance over standard models in both high- and low-prevalence settings.
• This explainable AI holds promise for improving breast MRI screening and clinical adoption.