Towards cross-center head and neck cancer detection: a multi-level domain alignment exploration.
Authors
Affiliations (10)
Affiliations (10)
- School of Life Sciences and Medical Engineering, Guangxi Medical University, Nanning, China.
- Center for Genomic and Personalized Medicine, Guangxi Key Laboratory for Genomic and Personalized Medicine, Guangxi Collaborative Innovation Center for Genomic and Personalized Medicine, Guangxi Medical University, Nanning, China.
- Guangdong Cardiovascular Institute, Guangdong Provincial People's Hospital (Guangdong Academy of Sciences), Southern Medical University, Guangzhou, China.
- Department of Radiology, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou, China.
- Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, Medical Research Center, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou, China.
- Department of Radiology, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
- Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application, Guangzhou, China.
- Medical Research Institute, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
- Nanning Research Institute, Guilin University of Electronic Technology, Nanning, China.
- Guangxi Academy of Artificial Intelligence, Nanning, China.
Abstract
Deep learning models for head and neck cancer (HNC) detection from computed tomography (CT) hold significant promise for improving early detection-a critical priority given that 5-year survival drops from 84% for localized disease to 39% for metastatic cases. However, robust cross-center deployment remains challenging because scanner vendors, acquisition protocols, reconstruction Q15 kernels, and patient populations vary across hospitals. To address this challenge, we propose MDA-Net, a multi-level domain generalization framework for cross-center 3D CT classification. The final MDA-Net configuration is deliberately designed as a training-time generalization strategy: it does not use target-domain batches, target-domain labels, target-domain statistics, or adaptive updates during inference. Instead, it performs a single forward pass with fixed parameters on each unseen target CT volume. Our framework integrates three complementary levels: (i) input-level inverse frequency weighted sampling, Fourier Domain Adaptation (FDA), and Mixup augmentation to mitigate class imbalance and diversify source styles; (ii) feature-level IBN-enhanced 3D ResNet-18, DANN, and CORAL to learn style-invariant and statistically aligned representations; and (iii) optimization-level Sharpness-Aware Minimization (SAM) and Stochastic Weight Averaging (SWA) to stabilize training under style perturbation. We evaluate MDA-Net on a multi-center dataset comprising 1,081 cases from three hospitals under a strict leave-one-center-out protocol. On the primary Foshan target domain, MDA-Net improves AUC from 0.5121 to 0.5869 and ACC from 0.4829 to 0.6000; reverse validation on another held-out center further yields AUC = 0.5887. Our stepwise ablation demonstrates that IBN provides the largest single-component gain, while FDA and CORAL become beneficial only when combined with SAM, SWA, Mixup, and a moderate adversarial weight of λ = 0.1. These results position MDA-Net as an exploratory yet practical, privacy-preserving, and computationally efficient domain generalization strategy for cross-center HNC detection.