Prognostic value of multimodal MRI tumor features in Glioblastoma multiforme using textural features analysis

Image-derived features (“radiomics”) are increasingly being considered for patient management in (neuro)oncology and radiotherapy. In Glioblastoma multiforme (GBM), simple features are often used by clinicians in clinical practice, such as the size of the tumor or the relative sizes of the necrosis and active tumor. First order statistics provide a limited characterization power because they do not incorporate spatial information and thus cannot differentiate patterns. In this work, we present the methodological framework for building a prognostic model based on heterogeneity textural features of multimodal MRI sequences (T1, T1-contrast, T2 and FLAIR) in GBM. The proposed workflow consists in i) registering the available 3D multimodal MR images and segmenting the tumor volume, ii) extracting image features such as heterogeneity metrics and shape indices, iii) building a prognostic model using Support Vector Machine by selecting, ranking and combining optimal features. We present preliminary results obtained for the classification of 40 patients into short (≤ 15 months) or long (> 15 months) overall survival, validated using leave-one-out cross-validation. Our results suggest that several textural features in each MR sequence have prognostic value in GBM, classification accuracy of 90% (sensitivity 85%, specificity 95%) being obtained by combining both T1 sequences. Future work will consist in i) adding more patients for validation using training and testing groups, ii) considering additional features, iii) building a fully multimodal MRI model by combining features from more than two sequences, iv) consider survival as a continuous variable and v) combine image-derived features with clinical and histopatholoigcal data to build an even more accurate model.