Spatial Multi-Task Learning for Breast Cancer Molecular Subtype Prediction from Single-Phase DCE-MRI

TL;DR

The paper addresses non-invasive molecular subtype prediction in breast cancer using single-phase DCE-MRI by introducing a spatial multi-task learning framework that jointly models ER, PR, HER2, and Ki-67. It leverages a shared encoder with a multi-scale spatial attention mechanism and ROI weighting to capture intratumoral and peritumoral heterogeneity in the absence of temporal kinetics. On 960 cases with external validation, the model achieves an average AUC of $0.858$ across biomarkers and $0.955$ for overall molecular subtype, outperforming radiomics and single-task baselines; ER/PR/HER2 individual AUCs reach $0.893$, $0.824$, and $0.857$, with Ki-67 MAE of $8.2\pm2.1$%. Ablation studies confirm the value of multi-task learning, multi-scale attention, and peritumoral features, while attention visualizations support interpretability and clinical trust.

Abstract

Accurate molecular subtype classification is essential for personalized breast cancer treatment, yet conventional immunohistochemical analysis relies on invasive biopsies and is prone to sampling bias. Although dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) enables non-invasive tumor characterization, clinical workflows typically acquire only single-phase post-contrast images to reduce scan time and contrast agent dose. In this study, we propose a spatial multi-task learning framework for breast cancer molecular subtype prediction from clinically practical single-phase DCE-MRI. The framework simultaneously predicts estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2) status, and the Ki-67 proliferation index -- biomarkers that collectively define molecular subtypes. The architecture integrates a deep feature extraction network with multi-scale spatial attention to capture intratumoral and peritumoral characteristics, together with a region-of-interest weighting module that emphasizes the tumor core, rim, and surrounding tissue. Multi-task learning exploits biological correlations among biomarkers through shared representations with task-specific prediction branches. Experiments on a dataset of 960 cases (886 internal cases split 7:1:2 for training/validation/testing, and 74 external cases evaluated via five-fold cross-validation) demonstrate that the proposed method achieves an AUC of 0.893, 0.824, and 0.857 for ER, PR, and HER2 classification, respectively, and a mean absolute error of 8.2\% for Ki-67 regression, significantly outperforming radiomics and single-task deep learning baselines. These results indicate the feasibility of accurate, non-invasive molecular subtype prediction using standard imaging protocols.

Figures (4)

• Figure 1: Spatio Multi-Task Learning Framework: Molecular Subtype Prediction from Single-Phase DCE-MRI
• Figure 2: Multi-perspective visualization of spatial attention mechanism on a median-performance case (ER AUC=0.893). (a) Original DCE-MRI slice (39/128). (b) Radiologist-annotated tumor ROI. (c) Learned attention heatmap (Dice=0.87 with ROI). (d) Attention overlay. (e) Spatial intensity profiles showing peak within ROI. (f) Statistical comparison across $N=74$ patients: 8.29$\times$ attention ratio ($p<0.001$).
• Figure 3: ROC curves comparing the proposed multi-task deep learning model against baseline classifiers for molecular subtype prediction.
• Figure 4: Radar chart comparing the proposed model against baseline classifiers across six performance metrics.