An initial experience of machine learning based on multi-sequence texture parameters in magnetic resonance imaging to differentiate glioblastoma from brain metastases.

Purpose: To evaluate the performance of a machine learning method based on texture parameters in conventional magnetic resonance imaging (MRI) in differentiating glioblastoma (GB) from brain metastases (METs).

Materials And Methods: In this retrospective study conducted between November 2008 and July 2017, we included 73 patients diagnosed with GB (n = 73) and METs (n = 53) who underwent contrast-enhanced 3 T brain MRI. Twelve histogram and texture parameters were assessed on T2-weighted images (T2WIs), apparent diffusion coefficient maps (ADCs), and contrast-enhanced T1-weighted images (CE-T1WIs). A prediction model was developed for a machine learning method, and the area under the receiver operating characteristic curve of this model was calculated through 5-fold cross-validation. Furthermore, machine learning method's performance was compared with three board-certified radiologists' judgments.

Results: Univariate logistic regression model showed that the area under the curve (AUC) was highest with the standard value of T2WIs (0.78), followed by the maximum value of T2WIs (0.764), minimum value of T2WIs (0.738), minimum values of CE-T1WIs and contrast of T2WIs (0.733), and mean value of T2WIs (0.724). AUC calculated using the support vector machine was comparable to that calculated by the three radiologists (0.92 vs. 0.72, p < .01; 0.92 vs. 0.73, p < .01; and 0.92 vs. 0.86, p = .096).

Conclusion: In differentiating GB from METs on the basis of texture parameters in MRI, the performance of the machine learning method based on convention MRI was superior to that of the univariate method, and comparable to that of the radiologists.