Deep Learning on Tumor Habitat Sub-regions vs. Whole-Tumor ROI: A Comparative Study of Fusion Strategies for Breast Cancer Molecular Subtype Prediction.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, The First Affiliated Hospital of Yangtze University, 8 Hangkong Road, Shashi District, Jingzhou City, Hubei Province, 434000, People's Republic of China.
- Department of Radiology, Zhuzhou Central Hospital, Tianyuan District, 116 South Changjiang Road, Zhuzhou, 412007, China.
- Department of Radiology, The First Affiliated Hospital of Yangtze University, 8 Hangkong Road, Shashi District, Jingzhou City, Hubei Province, 434000, People's Republic of China. [email protected].
- Department of Radiology, The First Affiliated Hospital of Yangtze University, 8 Hangkong Road, Shashi District, Jingzhou City, Hubei Province, 434000, People's Republic of China. [email protected].
Abstract
This study aimed to compare deep learning models based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) whole-tumor and habitat regions of interest (ROIs) for non-invasive preoperative luminal subtype prediction in invasive breast cancer, and to evaluate three multi-modal fusion strategies in terms of discrimination, calibration, and clinical utility. In this multicenter, retrospective study, 396 patients with invasive breast cancer were included. A total of 276 patients from one center were divided into training (n = 220) and validation (n = 56) cohorts, with 120 patients from an independent institution serving as the external test set. Two-dimensional maximum cross-sectional images of whole-tumor ROI (group A) and habitat sub-region ROIs (group B) were extracted from DCE-MRI. ResNet50-based models were trained with four fixed random seeds, and three fusion approaches were compared: image-level early fusion, decision-level weighted fusion, and LR-Stacking fusion. In the external test set, group B achieved a numerically higher area under the curve (AUC) than group A (0.8187 vs. 0.7901), while the two models showed complementary strengths in calibration and threshold-dependent metrics. Among fusion models, decision-level weighted fusion achieved the highest AUC (0.8341), the best calibration (Brier score 0.1421), and the highest net clinical benefit across most threshold ranges on decision curve analysis. DeLong tests showed no significant pairwise AUC differences (all P > 0.05). Decision-level weighted fusion outperformed other fusion strategies across all three evaluation dimensions, supporting its potential as a practical tool for non-invasive preoperative luminal subtyping of breast cancer.