Author: Huang Chi-Shiuan, Wu Chih-Chun, Hui-Yu Cathy Tsai, Chen Yan-Han, Chen Yi-Wei, Pan Yi-Ying 👨🔬
Affiliation: Institute of Nuclear Engineering and Science, National Tsing Hua University, Taipei Veterans General Hospital, Tri-Service General Hospital 🌍
Purpose:
This study aims to develop and validate a machine learning (ML) model based on MRI-derived radiomic features to predict progressive disease (PD) in glioblastoma (GBM) patients four months after Boron Neutron Capture Therapy (BNCT). Furthermore, the study seeks to identify key imaging biomarkers crucial for evaluating treatment response, with the goal of establishing quantitative imaging biomarkers.
Methods:
This retrospective study included 22 GBM patients treated with BNCT. Treatment responses were classified as PD (n = 15) or non-PD (n = 7) based on the Response Assessment in Neuro-Oncology (RANO) criteria. Radiomic features were extracted from tumor, necrosis, and edema regions in pre- and post-BNCT MRI scans. Delta-radiomics features defined as changes in feature values between pre- and post-BNCT MRIs, were calculated. Features selections was performed using the Wilcoxon–Mann–Whitney (WMW) test, followed by Max-Relevance and Min-Redundancy (mRMR) to identify the top 30 significant features. Six ML models including Logistic Regression (LR), K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Machine (SVM), Adaptive Boosting (AdaBoost), Gaussian Naive Bayes (GNB) were trained and validated with leave-one-out cross-validation (LOOCV). To address class imbalance, the synthetic minority oversampling technique (SMOTE) was applied.
Results:
Among the six machine learning models tested, the Logistic Regression (LR) model demonstrated the highest performance, achieving an accuracy of 100%, an area under the curve (AUC) of 1.00, sensitivity of 100%, and specificity of 100% in LOOCV. Other models also achieved high performance, with accuracy exceeding 80%. Additionally, radiomic features extracted from pre-BNCT MRI account for 50% of selected features.
Conclusion:
Radiomic features extracted from pre-BNCT MR images demonstrated strong predictive capability for GBM treatment response, with the LR model achieving optimal accuracy (100%). This ML-based approach offers valuable clinical potential in optimizing BNCT strategies for GBM, although further validation on larger cohorts is necessary to ensure broad applicability.