3712 - Radiomics-Based Prognosis in Nasopharyngeal Cancer: Preliminary Results in a Multicenter Study
Presenter(s)
S. Khongwirotphan1,2, J. Setakornnukul3, S. Chakrabandhu4, S. Chamchod5,6, C. Lowanichkiattikul7, A. Prayongrat8, C. Lertbutsayanukul9, S. Sriswasdi10,11, and Y. Rakvongthai2,12; 1Department of Radiological Technology and Medical Physics, Faculty of Allied Health Sciences, Chulalongkorn University, Bangkok, Thailand, 2Chulalongkorn University Biomedical Imaging Group, Department of Radiology, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand, 3Division of Radiation Oncology, Department of Radiology, Faculty of Medicine Siriraj Hospital, Mahidol University, Bangkok, Thailand, 4Faculty of Medicine, Chiang Mai University, Chiang Mai, Thailand, 5Princess Srisavangavadhana Faculty of Medicine, Chulabhorn Royal Academy, Bangkok, Thailand, 6Radiation Oncology Department, Chulabhorn Hospital, Chulabhorn Royal Academy, Bangkok, Thailand, 7Faculty of Medicine Ramathibodi Hospital, Mahidol University, Bangkok, Thailand, 8Division of Radiation Oncology, Department of Radiation, Faculty of Medicine, Chulalongkorn University, King Chulalongkorn Memorial Hospital, Bangkok, Thailand, 9Division of Radiation Oncology, Department of Radiology, Faculty of Medicine, Chulalongkorn University, King Chulalongkorn Memorial Hospital, Thai Red Cross Society, Bangkok, Thailand, 10Center for Artificial Intelligence in Medicine, Research Affairs, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand, 11Center of Excellence in Computational Molecular Biology, Chulalongkorn University, Bangkok, Thailand, 12Division of Nuclear Medicine, Department of Radiology, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand
Purpose/Objective(s): To construct radiomics-based prognostic models and evaluate their performance and generalizability through external test datasets of nasopharyngeal cancer (NPC) patients from four hospitals.
Materials/Methods: A total of 363 retrospective datasets of patients with NPC treated with radiotherapy were collected from four hospitals (183 from Hospital A as the training and validation set; and 50, 50, 35 and 45 from Hospitals A, B, C, D as the four external test sets). For each patient, radiomic features were computed from the gross tumor volume (GTV) in contrast-enhanced CT simulation image, while clinical features included age, sex, T-stage, N-stage, and overall stage. Based on 30 patients’ GTVs segmented independently from three radiation oncologists, the ratio between inter-operator and inter-patient standard deviations was calculated for each radiomic features and a threshold of 0.1 was applied to select 298 features that were robust to tumor delineation. For each of the survival outcomes, three prognostic models (clinical-only, radiomic-only, and combined clinical-radiomic) were constructed based on the Cox proportional hazard model with L2-regularization. For each initial input feature set (clinical-only, radiomic-only, or combined), recursive feature elimination (RFE) was first performed to remove uninformative features. Then, the L2-regularization strength was tuned to maximize the concordance index (C-index) on the validation set. During each RFE step and during hyperparameter tuning, 30 training-validation splits (2:1 patient ratio) were generated to statistically evaluate the distribution of model coefficients and C-index. Data augmentation was also performed by sampling new radiomic feature values based on the co-variance estimated from independent segmentations. The original clinical data were appended to the synthetic radiomic data.
Results: Combination of radiomic and clinical data (combined) significantly outperformed using only clinical data or radiomic data on the validation sets in all outcomes (Wilcoxon signed rank p-values < 0.05), achieving C-index of 0.65-0.66 (compared to 0.58-0.60 for clinical-only and 0.62-0.65 for radiomic-only). On four external test datasets, the combined model achieved C-index of 0.50-0.54 for overall survival (OS), 0.53-0.55 for progression-free survival (PFS), and 0.52-0.58 for distant metastasis-free survival (DMFS). Data augmentation, via 3-fold up-sampling of cases with early events within 3 years had a positive impact on the generalizability across all hospitals, with C-index of 0.55-0.57 for OS, 0.55-0.59 for PFS, and 0.54-0.61 for DMFS.
Conclusion: These preliminary results in multicenter cohorts showed that radiomics improved NPC prognostic performance as compared to clinical data and yielded moderate generalizability which could be improved through data augmentation.