Enhancing treatment response prediction in brain metastases: A multi-modal radiomics approach using MRI and CBCT

Purpose This study investigates a multi-modal radiomics approach that combines cone-beam computed tomography (CBCT) and magnetic resonance imaging (MRI) at both the image and feature levels to identify the optimal strategy for predicting treatment response in brain metastasis. Methods A retrospective analysis was conducted on 103 brain metastases treated between April 2019 to March 2024. Multi-modality approaches included image-level fusion and features-level combination (before and after selection). Single-modality models were developed using contrast-enhanced T1-weighted MRI, original CBCT, and resampled CBCT. Following image biomarker standardization initiative (IBSI) guidelines, 64-bin gray-level discretization with mean relative region-of-interest (ROI) ±3 SD intensity rescaling was applied, resulting in 716 extracted features. Features selection was performed using univariate analysis followed by correlation filtering to remove redundant features. Logistics regression with LASSO regularization and Support Vector Machine (SVM) classifiers were used for model development and comparison, with 10-fold cross-validation. Model performance was assessed using standard metrics: accuracy, sensitivity, specificity, and AUC. The Radiomic Quality Score (RQS) was also assessed. Results Of the 103 tumours, 67 demonstrated good treatment responses based on RANO-BM criteria. The MRI-only model achieved an accuracy of 72%, sensitivity of 85%, specificity of 48%, and AUC of 0.79[95% CI:0.61-0.93]. The original CBCT model yielded similar result (accuracy: 71%; AUC: 0.71, and 0.73, respectively). The best-performing model was the multi-modality combination of MRI and resampled CBCT features before selection (accuracy: 84%, sensitivity: 89%, specificity: 77%, AUC: 0.93[95%CI:0.71-1.00]). SVM models using the same feature sets produced comparable trends, confirming the robustness of the logistic regression findings. The RQS was 17/36 (47.2%). Conclusion Integrating radiomic features from both MRI and CBCT improves treatment response prediction in brain metastases and outperforms single-modality models. These findings underscore the potential of multi-modality radiomics to enhance personalised treatment planning. Further validation in larger, multi-centre cohorts is needed to confirm generalizability.