Non-Invasive Prediction of Glioma Tumor Stemness Using Multimodal MRI
Davoodi-Bojd E, Malta T, Griffith B, Snyder J, Walbert T, Noushmehr H, and Soltanian-Zadeh H. Non-Invasive Prediction of Glioma Tumor Stemness Using Multimodal MRI. Cancer Res 2019; 79(13).
Introduction : Stemness indices and sub-classifications of glioma have been proposed based on the epigenomic and transcriptomic makeup of these tumors, to inform prognosis and guide therapeutic management and drug discovery for this devastating disease. They reveal close association with tumor aggressiveness, which reflects clinical outcome. Stemness indices reveal tumor biology, which may affect sensitivity of individual tumors to therapy, and ultimately help to develop effective targeted therapies against this deadly disease. Due to the difficulty and cost of brain biopsy and molecular profiling, we proposed a noninvasive approach to estimate glioma stemness indices solely by numerical MRI features. Methods: We used four MRI modalities (pre- and post-contrast T1-weighted, T2-weighted, and T2-FLAIR images) of 73 glioma patients from the cancer genome atlas (TCGA) respiratory. For each patient, the tumor was segmented into four sub-volumes: necrosis, edema, contrast enhancing tumor (CE), and non-enhancing tumor (NE) using BraTumIA software. From each of these four sub-volumes, 25 features (1 volume, 4 histogram, and 20 texture features) were extracted from the four MRI modalities, generating a total of 400 noninvasive imaging features. A linear regression model was used to model each stemness index using the imaging features. Irrelevant/uncorrelated imaging features were discarded through a hierarchical feature selection-regression algorithm, which was developed to find the best subset of features predicting a particular stemness index linearly. The mean squared error (MSE) and Akaike information criterion (AIC) were used to optimize the regression models and to find an optimal number of features, respectively. Results: The resulting regression models were more accurate for the DNA-based stemness indices than the RNA-based stemness indices. This is not surprising, since it was reported that mDNAsi correlated with glioma outcome more than mRNAsi. Interestingly, ENHsi which defined stemness indices based on genomic enhancer stem signatures (targets of Sox and Oct transcription factor binding) had the strongest correlation with the MRI features. This suggests that MRI can detect glioma subtypes with strong stem like features, which were shown to coincide with more aggressive subtypes of brain cancer. Moreover, features extracted from post-contrast MRI contributed to the resulting models more than the other MRI modalities. This confirms the crucial role of contrastenhanced MRI in the diagnosis and treatment of brain glioma tumors. Conclusions: Stemness indices can be estimated using MRI features and used for glioma diagnosis, treatment planning, and prognosis instead of repeated brain biopsies and molecular profiling, which are invasive and costly.