RESEARCH presented at this year’s European Society of Radiology Annual Meeting (ECR) has demonstrated the potential of an artificial intelligence (AI) system in detecting breast cancer using abbreviated dynamic contrast-enhanced MRI (DCE-MRI). The system, which analyses high-risk screening and diagnostic MRI scans, was tested on datasets from five hospital groups across different countries and the publicly available Duke-Breast-Cancer-MRI dataset.
The AI model processes pre-contrast and a single post-contrast T1 image, identifying suspicious regions and assigning a malignancy score per breast on a scale from 1 to 10. To evaluate the system’s robustness, it was tested on screening data from the DENSE trial and hospital groups from Argentina, Switzerland, Turkey, and the United States.
The study reported strong performance, with the area under the receiver operating characteristic curve (AUROC) demonstrating high accuracy across diverse populations and imaging protocols. Results showed AUROCs of 0.891 (95% CI: 0.828-0.944) in Argentina (41 out of 780 exams contained biopsy-confirmed cancer), 0.863 (95% CI: 0.824-0.896) in Switzerland (98/3499), 0.955 (95% CI: 0.898-0.998) in Turkey (33/164), and 0.904 (95% CI: 0.877-0.929) in the United States (153/1096). The AI system exhibited consistent diagnostic performance across different scanners and imaging protocols.
Further analysis of the Duke-Breast-Cancer-MRI dataset, which exclusively contains cancer cases, involved a breast-level assessment, comparing cancerous and non-cancerous breasts. The AI system achieved an AUROC of 0.965 (95% CI: 0.957-0.972), comparable to earlier AI models using two post-contrast images.
To compare AI predictions with radiologist assessments, researchers evaluated exams with BI-RADS scores, finding moderate agreement (Cohen’s kappa = 0.502 (95% CI: 0.449-0.555)) between AI predictions (score ≥ 9) and radiologist interpretations (BI-RADS 1 or 2 vs. 4 or 5).
In screening-only data, the AI system maintained strong performance, with AUROCs of 0.812 (95% CI: 0.753-0.868) in a Dutch hospital (66/2920 exams with cancer) and 0.803 (95% CI: 0.747-0.856) in the DENSE trial dataset (83/517).
These findings highlight the potential of AI-assisted breast cancer detection in abbreviated MRI scans, offering decision-support capabilities for radiologists and potentially improving early cancer detection rates.
Reference
Eppenhof K et al. Evaluation of an AI System for Cancer Detection in Abbreviated Breast MRI. RPS 705. ECR Annual Meeting, 26 February–02 March, 2025.
