Detection and characterization of MRI breast lesions using deep learning

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Accurate differentiation of breast lesions is crucial for effective patient management.
• Magnetic Resonance Imaging (MRI) offers detailed anatomical and functional information for breast lesion assessment.
• Deep learning models demonstrate potential in medical image analysis, including breast lesion characterization.

Purpose of the Study:

• To evaluate a deep learning model's capability to differentiate benign from malignant breast lesions using MRI.
• To assess the model's ability to characterize various histological subtypes of breast lesions.
• To determine the performance of a deep learning approach in breast MRI analysis.

Main Methods:

• Development of a deep learning model for simultaneous lesion detection and characterization.
• Utilized single 2D T1-weighted fat-suppressed MR images post-gadolinium administration.
• Trained and validated the model on a dataset of 335 and 168 MR images, respectively, covering 17 histological subtypes.

Main Results:

• Achieved a weighted average receiver operating characteristic (ROC)-AUC of 0.817 on the training set via cross-validation.
• Attained a weighted mean AUC of 0.816 on an independent test set.
• Demonstrated robust performance in classifying breast lesions.

Conclusions:

• The supervised-attention deep learning model exhibits good performance for breast MRI analysis.
• The findings suggest the potential of this AI approach for clinical application.
• Further validation on a larger, independent cohort is necessary to confirm efficacy.