Brain metastases and BRAF About 40-60% of stage IV melanoma patients develop brain metastases. Of these, nearly 50% carry a BRAF mutation - most commonly BRAF V600E, which substitutes valine for glutamic acid. This mutation is the prerequisite for treatment with highly effective BRAF inhibitors such as vemurafenib and dabrafenib, which have demonstrated significant intracranial response rates.
The heterogeneity problem BRAF mutation status can differ between the primary tumor and brain metastases in up to 26% of patients. This means testing the primary tumor is insufficient - the brain lesion itself must be genotyped. However, most patients with brain metastases receive radiosurgery rather than surgery, so tissue for genetic testing is often unavailable.
Radiomics as a solution This study evaluated whether MRI radiomics - AI-powered quantitative feature extraction from routine brain MRI scans - could predict BRAF V600E mutation status non-invasively in melanoma brain metastases, potentially guiding targeted therapy decisions without surgery.
Patient population Fifty-nine patients with surgically resected melanoma brain metastases from two German university centers were included: 45 from Cologne (training/validation group) and 14 from Regensburg (independent test group). All underwent preoperative contrast-enhanced MRI and post-surgical genetic analysis confirming BRAF V600E status.
Feature extraction Using the PyRadiomics package, 1,316 radiomics features per MRI sequence were extracted from manually segmented tumor volumes - including 16 shape, 19 first-order, and 75 second-order texture features derived from gray level matrices. Features were calculated from original images and after wavelet and Laplacian of Gaussian filter transforms.
Robust feature selection and model training To avoid unreliable features, a test-retest analysis using image perturbation was performed, retaining only features with intraclass correlation coefficient (ICC) between 0.91-1.00. A 10-fold stratified cross-validation selected the top 100 features by mutual information. A linear support vector machine (SVM) was trained and then applied blinded to the independent Regensburg test cohort.
BRAF prevalence in cohorts In the training group (Cologne), 22 of 45 patients (49%) had BRAF V600E mutation. In the external test group (Regensburg), 8 of 14 (57%) were BRAF-mutated - similar prevalence reflecting real-world rates.
Radiomics model performance The final 6-parameter radiomics signature achieved an area under the ROC curve of 0.92 (sensitivity 83%, specificity 88%) on the external test data. This high diagnostic performance was achieved using only routinely acquired structural MRI sequences - no specialized imaging was required.
Clinical parameters alone are insufficient Patient age alone (Model 1) had weak predictive performance. The combination of clinical parameters and MRI radiomics features in the final model substantially outperformed clinical features alone, confirming that the imaging-derived quantitative features add meaningful diagnostic information beyond what is captured in patient demographics.
Guiding therapy when surgery is not planned For patients receiving radiosurgery as first-line brain metastasis treatment - the majority of cases - no tissue is available for molecular testing. The radiomics classifier could identify BRAF-mutated lesions and support the decision to add BRAF inhibitor therapy alongside radiosurgery.
Addressing tumor heterogeneity Even in patients who did undergo primary tumor testing, this tool addresses the clinically important scenario where the brain metastasis has acquired a different BRAF status through tumor evolution. Intracranial re-testing via radiomics enables genuinely patient-tailored therapy.
Reducing unnecessary invasive procedures If confirmed in larger prospective studies, this non-invasive BRAF prediction tool could reduce the need for stereotactic biopsies - minimally invasive but still procedure-associated with risks - in patients for whom therapy selection depends on BRAF status.
Small external validation cohort The external test set of only 14 patients is a significant limitation. Although performance was strong (AUC 0.92), larger multi-center prospective studies are needed to confirm these results and establish the model's reliability across different MRI scanners, field strengths, and patient populations.
Manual tumor segmentation Tumor masks were manually delineated by neuroradiologists - a time-consuming process that limits clinical scalability. Automated tumor segmentation using deep learning would be required before such a radiomics pipeline could be deployed in routine clinical workflows.
Future integration with other MRI sequences The study used only T1-weighted contrast-enhanced and T2-weighted sequences. Including pre-contrast T1, FLAIR, diffusion-weighted imaging, or perfusion imaging could further enrich the radiomics feature space and potentially improve predictive accuracy.