Transfer Learning Strategy Based on Unsupervised Learning and Ensemble Learning for Breast Cancer Molecular Subtype Prediction Using Dynamic Contrast‐Enhanced <scp>MRI</scp>

Abstract

BACKGROUND: Imaging-driven deep learning strategies focus on training from scratch and transfer learning. However, the performance of training from scratch is often impeded by the lack of large-scale labeled training data. Additionally, owing to the differences between source and target domains, analyzing medical image tasks satisfactorily via transfer learning based on ImageNet is difficult. PURPOSE: To investigate two transfer learning algorithms for breast cancer molecular subtype prediction (luminal and non-luminal) based on unsupervised pre-training and ensemble learning: M_EL and B_EL, using malignant and benign datasets as the source domain, respectively. STUDY TYPE: Retrospective. POPULATION: Eight hundred and thirty-three female patients with histologically confirmed breast lesions (567 benign and 266 malignant cases) were selected. In the 5-fold cross-validation, the malignant cohort was randomly divided into 5 subsets to form a training set (80%) and a validation set (20%). FIELD STRENGTH/SEQUENCE: 3.0 T, dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) using T1-weighted high-resolution isotropic volume examination. ASSESSMENT: First, three datasets acquired at different times post-contrast were preprocessed as unlabeled source domains. Second, three baseline networks corresponding to the different MRI post-contrast phases were built, optimized by a combination of mutual information maximization between high- and low-level representations and prior distribution constraints. Next, the pre-trained networks were fine-tuned on the labeled target domain. Finally, prediction results were integrated using weighted voting-based ensemble learning. STATISTICAL TESTS: Mean accuracy, precision, specificity, and area under receiver operating characteristic curve (AUC) were obtained with 5-fold cross-validation. P < 0.05 was considered to be statistically significant. RESULTS: Compared with a convolutional long short-term memory network, pre-trained VGG-16, VGG-19, and DenseNet-121 from ImageNet, M_EL and B_EL exhibited significantly more optimized prediction performance (specificity: 90.5% and 89.9%; accuracy: 82.6% and 81.1%; precision: 91.2% and 90.9%; AUC: 0.836 and 0.823, respectively). DATA CONCLUSION: Transfer learning based on unsupervised pre-training may facilitate automatic prediction of breast cancer molecular subtypes. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY: Stage 2.