Utility of Contrast-Enhanced MRI Radiomics Features combined with clinical indicators for Predicting Induction Chemotherapy Response in Primary Central Nervous System Lymphoma

Abstract

Abstract Purpose To evaluate the utility of contrast-enhanced magnetic resonance imaging (CE-MRI) radiomics features combined with clinical variables to predict induction chemotherapy (IC) response when treating primary central nervous system lymphoma (PCNSL). Methods A total of 131 patients with PCNSL (101 in the training set and 30 in the testing set) who had had contrast-enhanced MRI scans were retrospectively analyzed. Pyradiomics was used to extract radiomics features, and the clinical variables of the patients were collected. Radiomics prediction models were constructed using different combinations of feature selection methods and machine learning models, from which the best combination was selected. We screened clinical variables associated with treatment outcome and developed clinical prediction models. The prediction performance of radiomics model, clinical model, and combined model, which integrates the best radiomics model and clinical characteristics was independently assessed and compared using Receiver Operating Characteristic (ROC) curves. Results In total, we extracted 1598 features. The best radiomics model we chose used T-test and RFE for feature selection and logistic regression for model building. Serum Interleukin 2 Receptor (IL-2R) and ECOG Score were used to construct a clinical predictive model of induction chemotherapy response. The results of the test set showed that the combined prediction model (radiomics and IL-2R) had the highest area under the ROC curve at 0.868 (0.683,0.967), followed by the radiomics model at 0.857 (0.681,0.957), and the clinical prediction model (IL-2R and ECOG) at 0.618 (0.413,0.797). The combined model was significantly more accurate than the clinical model (AUC, 0.868 vs. 0.618, P < 0.05). While the radiomics model had slightly better predictive power than the clinical model, this difference was not statistically significant (AUC, 0.857 vs. 0.618, P > 0.05). Conclusions Our prediction model, which combines radiomics signatures from CE-MRI with serum IL-2R, can effectively stratify PCNSL patients before high-dose Methotrexate (HD-MTX) based chemotherapy.