Interpretable radiomics-based machine learning model for differentiating glioblastoma from primary central nervous system lymphoma using contrast-enhanced T1-weighted imaging

This study aimed to develop and validate an interpretable radiomics-based machine learning model using contrast-enhanced T1-weighted imaging (CE-T1WI) to differentiate glioblastoma (GB) from primary central nervous system lymphoma (PCNSL), while comparing the performance of models using high-order versus low-order features. A retrospective analysis was conducted on 383 patients with histopathologically confirmed diagnoses (226 GB cases with 226 samples; 157 PCNSL cases with 232 samples). Radiomic features were extracted from CE-T1WI sequences using PyRadiomics, including both low-order and high-order features. A sequential feature selection pipeline combining variance thresholding, minimum redundancy maximum relevance (mRMR), and least absolute shrinkage and selection operator (LASSO) was used to identify the most informative and stable radiomic features for model building. Ten machine learning algorithms, including LightGBM, logistic regression, and random forests, were utilized to construct classifiers. Model performance was evaluated based on area under the curve (AUC), accuracy, specificity, sensitivity, negative predictive value (NPV) and positive predictive value (PPV). A comparison of the average performance metrics across all ten models was conducted between the high-order and low-order feature models. Interpretability was provided through SHapley additive exPlanations (SHAP). Statistical analyses were conducted with SPSS version 25.0 and Python 3.10.16. The sum of 1316 high-order features were extracted, and after feature reduction and selection, 17 optimal features were retained for machine learning models. Additionally, 107 low-order features were reduced to 20 discriminative features. The models, particularly those based on high-order features, demonstrated exceptional diagnostic performance, with AUC values exceeding 0.95 in 9 out of 10 models in the test sets. Among the ten classifiers evaluated, the LGBM model emerged as the most robust performer, achieving a test set AUC of 0.955 and demonstrating the smallest discrepancy (0.001) between the training and test AUC values. High-order features significantly outperformed low-order features, with improvements in AUC, sensitivity, and NPV (p < 0.05). The SHAP provided an in-depth interpretation of the LGBM model's predictions, identifying key features such as original_firstorder_Kurtosis and exponential_GLDem_DependanceVariance as significant contributors, while offering both global and sample-specific perspectives. The study demonstrates the potential of using a CE-T1WI-derived radiomics approach combined with machine learning for distinguishing GB from PCNSL with high accuracy and interpretability. The model provides a practical, non-invasive diagnostic approach to support preoperative decision-making, particularly when biopsy or pathological sampling is challenging or uncertain. This approach has strong potential for clinical application in neuro-oncology.