Differentiating IDH status in human gliomas using machine learning and multiparametric MR/PET

Abstract

Background: The purpose of this study was to develop a voxel-wise clustering method of multiparametric magnetic resonance imaging (MRI) and 3,4-dihydroxy-6-[¹⁸F]-fluoro-L-phenylalanine (FDOPA) positron emission tomography (PET) images using an unsupervised, two-level clustering approach followed by support vector machine in order to classify the isocitrate dehydrogenase (IDH) status of gliomas.

Methods: Sixty-two treatment-naïve glioma patients who underwent FDOPA PET and MRI were retrospectively included. Contrast enhanced T1-weighted images, T2-weighted images, fluid-attenuated inversion recovery images, apparent diffusion coefficient maps, and relative cerebral blood volume maps, and FDOPA PET images were used for voxel-wise feature extraction. An unsupervised two-level clustering approach, including a self-organizing map followed by the K-means algorithm was used, and each class label was applied to the original images. The logarithmic ratio of labels in each class within tumor regions was applied to a support vector machine to differentiate IDH mutation status. The area under the curve (AUC) of receiver operating characteristic curves, accuracy, and F1-socore were calculated and used as metrics for performance.

Results: The associations of multiparametric imaging values in each cluster were successfully visualized. Multiparametric images with 16-class clustering revealed the highest classification performance to differentiate IDH status with the AUC, accuracy, and F1-score of 0.81, 0.76, and 0.76, respectively.

Conclusions: Machine learning using an unsupervised two-level clustering approach followed by a support vector machine classified the IDH mutation status of gliomas, and visualized voxel-wise features from multiparametric MRI and FDOPA PET images. Unsupervised clustered features may improve the understanding of prioritizing multiparametric imaging for classifying IDH status.

Keywords: 18F-DOPA PET; Clustering; Diffuse glioma; IDH mutation; MRI; Machine learning.

Fig. 1

Simplified graphical overview of the processing

Fig. 2

Patient selection flow-chart

Fig. 3

Component planes with SOM for each imaging parameter ranging from blue to red according to each value. Red means a high weight. The inter-class borderlines obtained by K-means clustering with K = 16 are shown on the SOM component planes as black lines between the nodes. Detailed profiles can be seen on the K-means clustering map (from label 1 to 16)

Fig. 4

Representative cases of IDH wild-type and mutant gliomas with 16-class clustering that shows the highest classification performance. The CE-T1WI, T2WI, and FLAIR image, and ADC, rCBV, and FDOPA maps are shown for each patient. Each color within the tumor ROIs corresponds to each label in the 16-color bar. The ratios of each label are shown in pie chart. The voxels of the label 3 (red) can be seen frequently for the IDH wild-type gliomas

Fig. 5

a) Box-whisker plots and b) radar charts of each label by 16-class clustering. a The box-whisker plots showing median and interquartile range for log-ratio values. * shows a significant difference. b Radar charts of six variables (CE-T1WI, T2WI, FLAIR, ADC, rCBV, and FDOPA PET) in each label by 16-class clustering