Improving radiomics-based differentiation of supratentorial malignant brain tumors preoperatively with diffusion-weighted imaging: A three-class machine learning algorithm

This is the authors' abstract. We don't add key points for this paper.

OBJECTIVES: To construct a three-class multiparametric radiomic model for classifying supratentorial brain tumors [high-grade glioma (HGG), brain metastases (BM), and primary lymphomas of the central nervous system (PCNSL)], and investigate whether diffusion-weighted imaging (DWI)-based radiomic features enhance diagnostic performance. METHODS: 364 subjects of Dataset 1 with supratentorial HGG (n = 150), BM (n = 126), and PCNSL (n = 88) were split into training (n = 222) and validation (n = 142) sets. 109 subjects of Dataset 2 were regarded as external validation set. In training set, a multiparametric radiomic model based on 31 selected features from conventional MRI (T1WI, T1c, T2WI, and FLAIR), apparent diffusion coefficient (ADC) maps, and a conventional radiomic model based on 28 selected features from conventional MRI were built using a three-class random forest method. The diagnostic performance of two models was assessed. The area under the receiver operating characteristic (ROC) curves (AUCs) of the two models were compared using Delong analysis. RESULTS: The AUCs of multiparametric radiomic model were 0.978, 0.918, and 0.914 for HGG, BM, and PCNSL, respectively, in internal validation set. And 0.916, 0.898, 0.883 in external validation set. The multiparametric radiomic model yielded significantly higher performance in classifying BM (P < 0.001, P = 0.001, internal and external validation sets) and PCNSL (P = 0.012, P = 0.001, internal and external validation sets). CONCLUSION: DWI-based radiomic features offered incremental value beyond conventional MRI and improved model performance in discriminating among the three common supratentorial malignant brain tumors (HGG, BM, and PCNSL).