2613 - A Distribution Adaptation-Based Framework for Overcoming Domain Shift in Multicenter CT Radiomics: Differentiating Tuberculosis from Fungal Pneumonia
Presenter(s)
L. Song1, H. Xiao2, G. Ren3, J. Cai4, and Y. Yin5; 1Shandong Cancer Hospital Affiliated to Shandong First Medical University, Jinan, Shandong, China, 2Shandong Provincial Key Medical and Health Laboratory of Pediatric Cancer Precision Radiotherapy (Shandong Cancer Hospital), Jinan, Shandong, China, 3Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, 4Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, Kowloon, Hong Kong, 5Department of Radiation Oncology Physics and Technology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China
Purpose/Objective(s): The clinical translation of quantitative CT radiomics is hindered by domain shift across different acquisition systems. We hypothesize that mathematically aligning heterogeneous feature spaces via Multicenter Distribution Adaptation (MDA) mitigates inter-scanner variability, thereby ensuring robust out-of-distribution generalizability. This framework is validated on a demanding clinical task: the non-invasive differential diagnosis of pulmonary tuberculosis (TB) and fungal pneumonia (FP)—a significant diagnostic challenge impacting treatment workflows in immunocompromised and oncology populations.
Materials/Methods: This retrospective multi-institutional study curated CT data from 528 pathologically confirmed patients (317 TB, 211 FP) across four independent centers. Following automated deep learning-based anatomical and lesion segmentation, 1,781 radiomic features were extracted per region of interest. To minimize site-specific noise, an optimized supervised-unsupervised feature selection strategy maximized cross-validation stability. The core MDA algorithm was applied to explicitly minimize marginal and conditional probability distribution discrepancies between centers using Maximum Mean Discrepancy. Generalizability was benchmarked against traditional machine learning classifiers (SVM, RF, XGBoost) using strict leave-one-center-out cross-validation protocols.
Results: The MDA framework successfully overcame scanner-induced inter-center variability, consistently outperforming conventional baseline models in all unseen external validation cohorts. When distinguishing TB from FP in data-sparse external validation settings (e.g., center C, n=178), the MDA model achieved an AUC of 0.914 for infection-specific regions, significantly surpassing the best traditional model (XGBoost, AUC=0.755; p<0.001). Visual confirmation using t-SNE demonstrated that prior to MDA, feature spaces clustered strongly by acquisition center, indicating severe domain shift. Post-adaptation, MDA successfully mapped features into a scanner-invariant subspace, yielding distributions distinctly separable by clinical pathology rather than institutional origin.
Conclusion: The MDA framework provides a robust, physics-informed solution to mitigate domain shift in multi-institutional CT radiomics. By successfully differentiating complex pulmonary infections across diverse datasets without performance degradation, this methodology clears a major technical hurdle. These findings have direct implications for clinical practice and research, demonstrating that distribution adaptation can standardize quantitative imaging biomarkers for reliable AI deployment in broader thoracic and radiation oncology applications.