Radiomics-Based Machine Learning in Differentiation Between Glioblastoma and Metastatic Brain Tumors

Purpose To investigative the diagnostic performance of radiomics-based machine learning in differentiating Glioblastomas (GBM) from metastatic brain tumors(MBTs) Method Current study involved in 134 patients diagnosed and treated in our institution between April 2014 and December 2018. Radiomics were extracted from contrast-enhanced T1 weight imaging (T1C). Thirty diagnostic models were built based on five selection methods and six classification algorisms. The sensitivity, specificity, accuracy, and area under curve(AUC) of each model were calculated, based on which optimal model was chosen. Result Two models represented promising diagnostic performance with AUC of 0.80. The first model was combination of Distance Correlation as selection method and Linear Discriminant Analysis(LDA) as classifier. In training group, the sensitivity, specificity, accuracy, and AUC were 0.75, 0.85, 0.80, and 0.80, respectively; and in testing group, the sensitivity, specificity, accuracy, and AUC of the model were 0.69, 0.86, 0.78, and 0.80, respectively. The second model was Distance Correlation as selection method and logistic regression(LR) as classifier, with sensitivity, specificity, accuracy, and AUC of 0.75, 0.85, 0.80, 0.80 in training group; and 0.69, 0.86, 0.78, 0.80 in testing group. Conclusion Radiomic-based machine-learning have potential to be utilized in differentiating GBM from MBTs.