Automatic Segmentation and Molecular Subtype Classification of Breast Cancer Using an MRI-based Deep Learning Framework

Abstract

Purpose: To build a deep learning framework using contrast-enhanced MRI for lesion segmentation and automatic molecular subtype classification ibreast cancer. Materials and Methods: This retrospective multicenter study included patients with biopsy-proven invasive breast cancer between January 2015 and Janua2021. An automatic breast lesion segmentation model was developed using three-dimensional (3D) ResU-Net as the backbone, and its accuracy was evaluated in an internal and two external testing datasets using the Dice score. An ensemble model for classification of breast cancer into four moleculasubtypes (Ensemble ResNet) was then developed by combining both two-dimensional and 3D lesion features. The performance of Ensemble ResNet wevaluated in the three testing datasets using the area under the receiver operating characteristic curve (AUC). Results: A total of 687 female patients (mean age ± SD, 48.70 years ± 8.97) were included, with 289, 61, 73, and 264 patients included in the training, internal testing, and two external testing datasets, respectively. The proposed segmentation model achieved high accuracy in internal testing dataset 1, external testing dataset 2, and external testing dataset 3 (Dice scores: 0.86, 0.82, 0.85) and luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, and triple-negative breast cancer (TNBC) subtypes (Dice scores: 0.8571, 0.8323, 0.8199, 0.8481). Ensemble ResNet demonstrated high performance for the prediction of luminal A subtypes (AUC range, 0.74–0.84), luminal B subtypes (AUC range, 0.68–0.72), HER2-enriched sub-types (AUC range, 0.73–0.82), and TNBC (AUC range, 0.80–0.81) in the three testing datasets. Conclusion: The proposed novel deep learning framework based on MRI achieved high, robust performance in fully automatic classification of breast camolecular subtypes.