Related Experiment Video
Updated: Aug 14, 2026
15:48
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
Artificial Intelligence for Breast MRI Lesion Classification: A Targeted Evidence Synthesis and Meta-Analysis of
Romuald Ferre1, Thad Benefield2, Cherie M Kuzmiak1
1Division of Breast Imaging, Department of Radiology, UNC School of Medicine, Chapel Hill, NC 27599, USA.
Diagnostics (Basel, Switzerland)
|August 13, 2026
Summary
Artificial intelligence (AI) shows promise in classifying breast lesions on MRI scans, achieving a pooled area under the curve (AUC) of 0.898. However, significant heterogeneity and limited validation require cautious interpretation for clinical use.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- Contrast-enhanced breast MRI is crucial for breast lesion classification.
- Artificial intelligence (AI) methods are increasingly applied to medical imaging analysis.
- Evaluating the diagnostic performance of AI for breast lesion classification is essential.
Purpose of the Study:
- To synthesize the diagnostic performance of AI methods for classifying breast lesions on contrast-enhanced breast MRI.
- To estimate the pooled area under the receiver operating characteristic curve (AUC) for AI classification of breast lesions.
- To assess the generalizability and reliability of AI models in breast lesion diagnosis.
Main Methods:
- A targeted evidence synthesis and meta-analysis adhering to PRISMA 2020 principles.
- Inclusion of 9 studies applying machine learning or deep learning to contrast-enhanced breast MRI for lesion classification.
- Pooling of AUCs using a random-effects model with heterogeneity quantified by I-squared.
Main Results:
- The pooled random-effects AUC for AI classification of breast lesions was 0.898 (95% CI, 0.875-0.918).
- Substantial heterogeneity (I-squared = 88.1%) was observed, with a 95% prediction interval of 0.824-0.943.
- Study evaluation set sizes ranged from 60 to 3936 participants.
Conclusions:
- AI models demonstrate promising discriminative performance for breast lesion classification on contrast-enhanced MRI within the studied corpus.
- Significant heterogeneity, variations in analysis units, and limited external validation temper confidence in the generalizability of AI models.
- Future research should focus on multi-institutional external validation, transparent reporting, and prospective evaluations before routine clinical deployment of AI in breast MRI analysis.