Machine learning in preoperative glioma MRI: Survival associations by perfusion‐based support vector machine outperforms traditional MRI

Abstract

PURPOSE: To retrospectively evaluate the performance of an automatic support vector machine (SVM) routine in combination with perfusion-based dynamic susceptibility contrast magnetic resonance imaging (DSC-MRI) for preoperative survival associations in patients with gliomas and compare our results to traditional MRI. MATERIALS AND METHODS: The study was approved by the Ethics Committee and informed consent was signed. Structural, diffusion- and perfusion-weighted MRI was performed at 1.5-T preoperatively in 94 adult patients (49 males, 45 females, 23-82 years; mean 51 years) later diagnosed with a primary glioma. Patients were randomly assigned in training and test datasets and the resulting DSC-based survival associations by SVM were compared to traditional MRI features including contrast-agent enhancement, perfusion- and diffusion-weighted imaging, tumor size, and location. The results were adjusted for age, neurological status, and postoperative factors associated with survival, including surgery and adjuvant therapy. RESULTS: For 1- (26/33 alive, 11/14 deceased), 2- (15/21, 21/26), 3- (12/16, 27/31) and 4- (12/15, 28/32) year survival associations in the test dataset (47 patients), the SVM routine was the only biomarker to consistently associate with survival (Cox; P < 0.001). CONCLUSION: The automatic machine learning routine presented in our study may provide the operator with a reliable instrument for assessing survival in patients with glioma.