Automated full-process pipeline via multi-parametric MRI for tumor segmentation, molecular subtype classification and prognostic factor analysis in breast cancer.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Department of Computer Center, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Department of Thyroid and Breast Surgery, The Central Hospital of Wuhan, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Department of Thyroid and Breast Surgery, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Department of Radiology, The University of Hong Kong-Shenzhen Hospital, Shenzhen, China.
- Department of Radiology, Xiangyang Central Hospital, Affiliated Hospital of Hubei University of Arts and Science, Xiangyang, China.
Abstract
Accurate identification of molecular subtypes and prognostic factors is crucial for personalized breast cancer management. Multi-parametric magnetic resonance imaging (mp-MRI) offers comprehensive, non-invasive characterization of tumors by capturing complementary information on morphology, perfusion, and tissue microstructure, but conventional radiologic assessment is often inadequate for reliably inferring molecular subtype or prognostic factors. Existing deep learning (DL) approaches depend on manually delineated tumor regions to define inputs, which is difficult to scale in routine practice. We therefore propose an automated end-to-end DL pipeline that jointly performs tumor segmentation and biomarker prediction directly from mp-MRI. In this multicenter study, mp-MRI data from 484 patients with breast cancer were collected from three institutions. The pipeline first used a three-dimensional U-Net (3D U-Net) for automated tumor segmentation and then a two-dimensional Residual Network (2D ResNet)-based classifier for predicting molecular subtypes and prognostic factors. Segmentation performance was quantified using mean intersection over union (mIoU) and mean dice similarity coefficient (mDice). Classification performance was assessed through area under the receiver operating characteristic curve (AUC), complemented by accuracy, sensitivity, specificity, and F1-score metrics. The DeLong test was used to compare AUCs between models based on automated segmentation regions of interest (ROIs) and manually delineated ROIs. The model achieved robust segmentation performance in the validation set (mIoU =0.772, mDice =0.864) and external cohorts (mIoU =0.732-0.746, mDice =0.838-0.848). For molecular subtyping, triple-negative (TN) differentiation achieved the best performance (AUC: validation =0.839, external =0.729-0.749). Axillary lymph node (ALN) status prediction yielded AUCs of 0.773, 0.715, and 0.700, respectively. Ki67 expression prediction demonstrated high generalizability (AUC: validation =0.788, external =0.749-0.784). No statistically significant differences were observed in classification performance between models based on automated segmentation ROIs and those based on manually delineated ROIs (all P>0.05). The automated end-to-end pipeline integrates 3D U-Net-based tumor segmentation and 2D ResNet-based biomarker classification, supporting its potential as a complementary tool for diagnostics, treatment planning, and prognostic assessment in breast cancer.