Impact of ComBat harmonization on MRI radiomic-analysis of nasopharyngeal carcinoma: a dual-center study.

Purpose: To investigate the impact of ComBat harmonization on the performance of 1.5T/3.0T MRI radiomic features in predicting the overall survival(OS) of patients with locally advanced nasopharyngeal carcinoma(LANPC). Methods: This dual-center retrospective study included 573 patients with LANPC (435 male and 138 female patients). The 3.0T MRI dataset from Hospital 1 was used as the reference batch (training cohort, n=287), and the 1.5T MRI dataset from Hospital 2 was subjected to ComBat harmonization. The 1.5T MRI dataset before ComBat harmonization served as Validation Cohort 1 (n=286), while the 1.5T MRI dataset after ComBat harmonization was designated as Validation Cohort 2. Radiomics features were extracted from the segmented tumors, and principal component analysis (PCA) was applied to assess the batch effect of radiomics features between the two validation cohorts. A radiomics-clinical prognostic model was developed using radiomic features and other clinical factors using multivariate Cox regression. The concordance index (C-index) was used to evaluate model performance in predicting OS, and Kaplan-Meier survival analysis was conducted to explore the impact of the model on the prognosis of LANPC patients. Results: PCA without ComBat revealed noticeable differences in the first two principal components between batches, indicating a batch effect or unstable radiomic features. Following ComBat harmonization, the principal components showed more consistency between batches, demonstrating radiomics feature stability between batches. Multivariate Cox regression identified EBV-DNA and platelet count as independent clinical factors, which were integrated with the radiomics score to construct the final prognostic model. The C-indexes of the model for predicting OS in the training, validation 1 and validation 2 cohorts was 0.797, 0.610 and 0.648, respectively. The 5-year OS for the model defined low-risk group was significantly better than that of the high-risk group (P < 0.001). Conclusion: ComBat harmonization effectively reduced the inter-batch effect of radiomic feature sets across different centers and scanners. While the model maintained favorable discrimination for OS and robust risk-group separation via Kaplan-Meier analysis, the primary benefit of ComBat harmonization was the improved qualitative consistency of risk stratification between multi-center datasets, which may support the clinical application of risk-adapted therapy for LANPC.