Abstract
Background: The accurate and early distinction of glioblastomas (GBMs) from single brain metastases (BMs) provides a window of opportunity for reframing treatment strategies enabling optimal and timely therapeutic interventions. We sought to leverage physiologically sensitive parameters derived from diffusion tensor imaging (DTI) and dynamic susceptibility contrast (DSC)-perfusion-weighted imaging (PWI) along with machine learning-based methods to distinguish GBMs from single BMs. Methods: Patients with histopathology-confirmed GBMs (n = 62) and BMs (n = 26) and exhibiting contrast-enhancing regions (CERs) underwent 3T anatomical imaging, DTI and DSC-PWI prior to treatment. Median values of mean diffusivity (MD), fractional anisotropy, linear, planar and spheric anisotropic coefficients, and relative cerebral blood volume (rCBV) and maximum rCBV values were measured from CERs and immediate peritumor regions. Data normalization and scaling were performed. In the next step, most relevant features were extracted (non-interacting features), which were subsequently used to generate a set of new, innovative, high-order features (interacting features) using a feature engineering method. Finally, 10 machine learning classifiers were employed in distinguishing GBMs and BMs. Cross-validation and receiver operating characteristic (ROC) curve analyses were performed to determine the diagnostic performance. Results: A random forest classifier with ANOVA F-value feature selection algorithm using both interacting and non-interacting features provided the best diagnostic performance in distinguishing GBMs from BMs with an area under the ROC curve of 92.67%, a classification accuracy of 87.8%, a sensitivity of 73.64% and a specificity of 97.5%. Conclusions: A machine learning based approach involving the combined use of interacting and non-interacting physiological MRI parameters shows promise to differentiate between GBMs and BMs with high accuracy.