Abstract
BACKGROUND: The present study aimed to establish a robust predictive model based on a machine learning (ML) algorithm providing preoperative noninvasive diagnosis and to further explore the contribution of each magnetic resonance imaging (MRI) sequence to the classification to help select images for future model development. METHODS: This was a retrospective cross-sectional study, and consecutive patients with histologically confirmed diffuse gliomas in our hospital from November 2015 to October 2019 were recruited. The participants were grouped into a training and testing set based on a ratio of 8:2. Five MRI sequences were employed to develop the support vector machine (SVM) classification model. An advanced contrast analysis of single-sequence-based classifiers was performed, according to which different sequence combinations were tested, and the best one was selected to form an ultimate classifier. Patients whose MRIs were acquired with other types of scanners formed an additional, independent validation set. RESULTS: A total of 150 patients with gliomas were used in the present study. Contrast analysis revealed that the contribution of the apparent diffusion coefficient (ADC) was the most significant [accuracies were as follows: histological phenotype, 0.640; isocitrate dehydrogenase (IDH) status, 0.656; and Ki-67 expression, 0.699] and that of T1 weighted imaging was limited (accuracies were as follows: histological phenotype, 0.521; IDH status, 0.492; and Ki-67 expression, 0.556). The ultimate classifiers for IDH status, histological phenotype, and Ki-67 expression achieved promising performances with area under the curve (AUC) values of 0.88, 0.93, and 0.93, respectively. The classifiers for the histological phenotype, IDH status, and Ki-67 expression correctly predicted 3 of 5 subjects, 6 of 7 subjects, and 9 of 13 subjects in the additional validation set, respectively. CONCLUSIONS: The present study showed satisfactory performance in predicting the IDH genotype, histological phenotype, and Ki-67 expression level. The contrast analysis revealed the contribution of different MRI sequences and suggested that the combination of all the acquired sequences was not the optimal strategy to build the radiogenomics-based classifier.