Automatic Differentiation between Pseudoprogression and Tumor Recurrence by Applying an SVM to Multiparametric MRI-Data

Objective: The aim of this retrospective study was to evaluate whether a differentiation of pseudoprogression and tumor recurrence in patients with resected glioblastoma multiforme was possible by applying a machine learning approach to data obtained by several different magnetic resonance imaging (MRI) sequences.

Methods: Data from 8 patients with pseudoprogression and 8 patients with tumor recurrence was used in this retrospective study. For each patient 8 images, obtained by multiple MRI sequences, were registered, such that each voxel was characterized by an 8-dimensional feature vector. A region of interest (ROI) was drawn over the contrastenhanced lesion in the T1-weighted image. A one-class support vector machine (OCSVM) was trained on the feature vectors of the voxels within the ROI of patients with pseudoprogression. The classifier was tested using cross-validation. The percentage of voxels within the ROI that were classified by the SVM to represent pseudoprogression was used to make a decision whether a patient suffered from recurrent tumor or showed pseudoprogression.

Results: The single voxels were classified with an area under the ROC-curve of 0.66. The percentages of voxels that were thought to represent pseudoprogression were significantly larger with a p-value of 0.0104 in patients with pseudoprogression compared to patients with tumor progression. The sensitivity and specificity with which the single patients were classified were 0.75 and 0.875 respectively.

Conclusion: The results showed that a differentiation based on MRI data using a machine learning approach is possible. However the excellent results of previous research by Hu et al. [1] could not be reproduced. Especially adjustments to the MRI protocols are expected to improve the method.