OBJECTIVES: This study aims to develop and validate a novel multimodal interpretable artificial intelligence model capable of fusing radiomics features and imaging features to accurately classify primary central nervous system lymphoma (PCNSL) and glioblastoma (GBM), and to explore the impact of including the tumor and different peritumoral brain zones (PBZs) on the diagnostic performance of the model. MATERIALS AND METHODS: A retrospective cohort of 242 patients with PCNSL or GBM, along with their multi-sequence MRI scans and clinical data, was enrolled from two medical centers. After image preprocessing, tumor and PBZ regions were segmented as volumes of interest (VOIs), and imaging features were evaluated. Based on different combinations of three image sequences and four VOIs, a total of seven models (T1WI-Tumor-Model, T2WI-Tumor-Model, T1CE-Tumor-Model, T1CE-Tumor-Edema-Model, T1CE-Tumor-PBZ10mm-Model, T1CE-Tumor-PBZ20mm-Model, and Multimodal-Radiomics-Model) were established. Receiver operating characteristic (ROC) curves, sensitivity, specificity, calibration curves, and Decision Curve Analysis (DCA) were used to assess the models' performance and clinical application value. Additionally, the best model was analyzed for interpretability using SHapley Additive exPlanations (SHAP). RESULTS: The T1CE-Tumor-Edema-Model and Multimodal-Radiomics-Model demonstrated superior predictive performance among the seven models, with AUCs of 0.91 and 0.94 in the external validation cohort, respectively. The SHAP interpretation of the results revealed that the Rad-Score and Clinical-Rad-Score features contributed the most to the models' decision-making process. CONCLUSION: A radiomics framework based on interpretable machine learning is proposed. Based on this framework, models combining different ranges of peritumoral brain regions with tumors were constructed for the first time, and radiomics features were fused with imaging features to improve classification accuracy and interpretability in distinguishing between PCNSL and GBM.
Zhao et al. (Tue,) studied this question.