Cechy radiomiczne z rezonansu magnetycznego przewidują status mutacji BRAF V600E w ganglioglasami

Abstract Background Ganglioglioma (GG) is one of the most common pathological types of drug-resistant epilepsy, yet its preoperative diagnosis remains challenging. The BRAF V600E mutation is the most frequent genetic alteration in GGs and has important reference value for the preoperative diagnosis and treatment strategy planning of this disease. Currently, BRAF V600E mutation status relies on surgical or biopsy tissue, highlighting the urgent need for a non-invasive preoperative prediction method to guide diagnostic and therapeutic decision-making. Methods We retrospectively included 114 patients with GG and known BRAF V600E mutation status. Preoperative T2-FLAIR images were normalized and resampled. Thirty cases were randomly selected for volume-of-interest (VOI) delineation by two physicians, and radiomic features with an intraclass correlation coefficient > 0.75 were retained. Patients were split 3:1 into a training set (n = 85) and a validation set (n = 29). After Z-score normalization and Spearman correlation analysis (ρ = 0.85) for redundancy reduction in the training set, recursive feature elimination (RFE) was applied to select the optimal feature subset. Five models—logistic regression (LR), support vector machine (SVM), ridge regression (Ridge-LR), linear discriminant analysis (LDA), and Naive Bayes—were then constructed and evaluated using AUC and decision curve analysis on both the training and validation sets. Results Among the 114 patients, 70 (61.4%) were BRAF V600E-mutant. No significant differences were observed between the BRAF V600E-mutant and wild-type groups in terms of age, sex, tumor location, disease duration, or age at epilepsy onset (all p > 0.05). However, the CD34 positivity rate was significantly higher in the mutant group than in the wild-type group (p = 0.001). In model validation, logistic regression (LR) achieved the best performance, with an AUC of 0.763 (95% CI: 0.555–0.932), accuracy of 75.9%, sensitivity of 88.9%, and specificity of 54.5% on the validation set. Decision curve analysis showed that LR, LDA, Ridge-LR, and SVM all yielded positive net clinical benefit within a threshold probability range of 0–0.65, with LR and LDA performing best in the 0.1–0.4 threshold range. Conclusion Routine preoperative MRI-based radiomics can non-invasively predict BRAF V600E mutation in GGs with good accuracy, with logistic regression performing best. This approach may provide molecular-level decision support for epilepsy surgery planning.