Development and external validation of an MRI radiomics-based machine learning model to predict tumour recurrence in brain metastases treated with Gamma Knife radiosurgery.
Authors
Affiliations (5)
Affiliations (5)
- Diagnostic Imaging and Radiotherapy, CODTIS, Faculty of Health Sciences, National University of Malaysia, Jalan Raja Muda Aziz 50300 Kuala Lumpur, Malaysia; Department of Medical Radiography, Ahmadu Bello University, Zaria, Nigeria.
- Department of Radiology, Faculty of Medicine, Universiti Kebangsaan Malaysia, Jalan Yaacob Latif, Bandar Tun Razak, Kuala Lumpur Malaysia.
- Gamma Knife Centre, Faculty of Medicine Universiti Kebangsaan Malaysia, Kuala Lumpur, Malaysia.
- Photonics Research Center, University of Malaya, 50603 Kuala Lumpur, Malaysia.
- Diagnostic Imaging and Radiotherapy, CODTIS, Faculty of Health Sciences, National University of Malaysia, Jalan Raja Muda Aziz 50300 Kuala Lumpur, Malaysia. Electronic address: [email protected].
Abstract
Predicting recurrence after gamma knife radiosurgery (GKRS) is clinically important, as it informs salvage treatment and patient management. MRI-based radiomics combined with machine learning (ML) offers promise for predicting tumour recurrence in brain metastasis, yet most studies lack external validation, limiting clinical translation. This study developed and externally validated radiomics-based models for predicting recurrence in brain metastases treated with GKRS. The primary dataset comprised 103 metastases from our institution (23 recurrent) and 125 lesions from the Cancer Imaging Archive (TCIA; 20 recurrent) for external validation. IBSI-compliant radiomics features were extracted from contrast-enhanced T1-weighted MRI using 64-bin grey-level discretisation and mean relative ROI ± 3 SD intensity rescaling, with and without LOG filtering. Feature selection combined correlation analysis and LASSO regression. Logistic regression classifiers were trained on 80 % of the data and tested on 20 %, followed by external validation. Five models were developed: MRI-only, LOG-filtered MRI, MRI + clinical, LOG-filtered MRI + clinical, and clinical + dosimetric. SHAP analysis was used for feature attribution, and methodological rigour was assessed using the Radiomics Quality Score (RQS). The best-performing model (LOG-filtered MRI + clinical features) achieved 81 % accuracy and an AUC of 0.93 in internal testing, and 79 % accuracy with an AUC of 0.78 on external validation, demonstrating strong robustness and generalisability. Adding clinical features significantly improved performance compared with MRI-only models. SHAP analysis revealed that tumour shape complexity (Compactness2 [IBSI: BQWJ]) and voxel intensity heterogeneity (IntensityRange[IBSI: 2OJQ]) were strong predictors, while maximum dose, gender, and primary tumour site also contributed among the clinical factors. The MRI-based radiomics model integrating LOG-filtered MRI features with clinical variables achieved high external validation performance. Unlike many black-box models that prioritise accuracy with limited interpretability, this approach combines predictive strength with transparent feature attribution, enhancing biology interpretability and supporting clinical translation. Prospective multi-centre studies are warranted to confirm its clinical utility.