195 - Evaluation of ComBat Batch Correction for Harmonizing Radiomic Features Across a Multicenter DCE-MRI Dataset

Purpose: To investigate the presence and impact of batch effects in radiomic features extracted from hepatocyte-avid DCE-MRI of metastatic ulveal melanoma lesions, and to evaluate the effectiveness of ComBat batch correction using image acquisition covariates such as the imager manufacturer. Batch effects, introduced by site-specific imaging protocols, scanner hardware, and contrast agents, can obscure biological signals and reduce the generalizability of machine learning models. Therefore, robust batch effect correction is essential to ensure that machine learning models capture clinically meaningful relationships rather than site-specific or covariate artifacts.

Materials and Methods: We retrospectively analyzed DCE-MRI studies from 401 patients. Each scan consisted of a T1-weighted hepatocyte-avid axial sequence acquired at multiple sites before IR treatment. An expert radiologist annotated a total of 1,537 liver metastases. Of these, 723 lesions from 157 patients were segmented using a fuzzy C-means algorithm. Morphological, texture, and kinetic radiomic features were subsequently extracted from each lesion. After segmentation quality control was performed using a DICE threshold of 0.25, 451 lesions from 136 patients were retained. Dimensionality reduction techniques (PCA, t-SNE, UMAP) were employed to visualize potential batch effects. Cosine similarity was used to quantify changes in feature similarity. The dataset was split 70/30 into training and testing subsets. ComBat was fit on the training set, and the testing set was corrected using the training-derived estimates.

Results: Dimensionality reduction revealed significant batch effects driven by scanner manufacturer and contrast agent, with intensity-based features showing greater susceptibility than morphological features. Following the post-ComBat correction, within-group similarity increased across most covariates, indicating improved internal consistency. However, between-group similarity also rose in some cases, notably among responders in the manufacturer subgroup, suggesting incomplete separation.

Conclusion: Radiomic features extracted from DCE-MRI are affected by site-specific acquisition parameters, introducing confounding batch effects. ComBat correction improved within-group consistency but did not fully resolve between-group separation, especially in imbalanced subgroups. These findings underscore the importance of covariate-aware batch correction in radiomics studies, suggesting that further refinement of feature harmonization strategies may be necessary.