Click here to


Are you sure ?

Yes, do it No, cancel

Use of Deep Learning in the Classification of Benign Lesions, Luminal A Cancers, and Other Molecular Cancer Subtypes in Breast Magnetic Resonance Imaging

H Whitney1,2*, N Antropova2 , M Giger2 , (1) Wheaton College, Wheaton, IL, (2) University of Chicago, Chicago, IL


(Sunday, 7/29/2018) 3:30 PM - 4:00 PM

Room: Exhibit Hall | Forum 8

Purpose: Transfer learning with convolutional neural networks (CNN) has been demonstrated for classification of benign lesions and malignant cancers using magnetic resonance (MR) breast images. Here we aim to investigate the performance of such techniques in classification of benign lesions and cancers by molecular subtypes in three tasks: (1) between benign lesions and luminal A cancers, (2) between benign lesions and other molecular subtypes (luminal B, basal-like, normal-like, HER2-enriched), and (3) between luminal A cancers and the other molecular subtypes.

Methods: The database of this HIPAA-compliant, IRB-approved retrospective study consisted of 194 benign lesions, 269 luminal A cancers, and 138 cancers of other molecular subtypes. For each case, a central slice from dynamic contrast enhanced (DCE) MR imaging was selected at the second post-contrast time point. For each lesion, an inclusive region of interest, which varied between 1 and 1.5 times the lesion size, was identified. For each classification task, CNN analysis was conducted using the VGG-19 net, pre-trained using ImageNet. Classification using features extracted from five max-pool layers was accomplished using a support vector machine with Gaussian radial basis function kernel and ten-fold cross validation; the hyperparameters were optimized using Bayesian optimization. The resulting posterior probability of cancer for the lesions was used to evaluate classification performance with receiver operating characteristic (ROC) curve analysis using the conventional binormal model. Area under the ROC curve (AUC) served as figure of merit.

Results: Classification of benign lesions and luminal A cancers, benign lesions and other subtypes, and luminal A cancers and the other subtypes yielded AUCs of 0.85±0.02, 0.90±0.02, and 0.67±0.03, respectively.

Conclusion: This work demonstrates that pre-trained CNN and transfer learning methods have strong performance in classification of benign lesions and molecular subtypes of breast cancer, and thus potential application in precision medicine.

Funding Support, Disclosures, and Conflict of Interest: M. Giger is a stockholder in R2/Hologic, co-founder and equity holder in Quantitative Insights, and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba.


Feature Extraction, Breast, Quantitative Imaging



Contact Email