A Neural Network Approach to Identify Glioblastoma Progression Phenotype from Multimodal MRI

SIMPLE SUMMARY: Glioblastoma is the most common malignant primary brain tumor and has a poor prognosis with inevitable recurrence or progression. The phenotypes of its progression patterns can be diverse, which may potentially affect the treatment plan and clinical outcome. Our study aimed to identify its progression pattern before surgery by using multimodal MRI. The results showed the different progression phenotypes are clinically important, and by using quantitative MR radiomics, together with neural network-based imaging analysis, we can predict glioblastoma progression phenotypes preoperatively. ABSTRACT: The phenotypes of glioblastoma (GBM) progression after treatment are heterogeneous in both imaging and clinical prognosis. This study aims to apply radiomics and neural network analysis to preoperative multimodal MRI data to characterize tumor progression phenotypes. We retrospectively reviewed 41 patients with newly diagnosed cerebral GBM from 2009–2016 who comprised the machine learning training group, and prospectively included 18 patients from 2017–2018 for data validation. Preoperative MRI examinations included structural MRI, diffusion tensor imaging, and perfusion MRI. Tumor progression patterns were categorized as diffuse or localized. A supervised machine learning model and neural network-based models (VGG16 and ResNet50) were used to establish the prediction model of the pattern of progression. The diffuse progression pattern showed a significantly worse prognosis regarding overall survival (p = 0.032). A total of 153 of the 841 radiomic features were used to classify progression patterns using different machine learning models with an overall accuracy of 81% (range: 77.5–82.5%, AUC = 0.83–0.89). Further application of the pretrained ResNet50 and VGG 16 neural network models demonstrated an overall accuracy of 93.1 and 96.1%. The progression patterns of GBM are an important prognostic factor and can potentially be predicted by combining multimodal MR radiomics with machine learning.