Abstract
Multimodal MRI is used for the evaluation of tumor burden in glioblastoma (GBM) patients after surgical resection, radiation, and chemotherapy. However, it can be challenging to identify non-enhancing infiltrative tumor and differentiate recurrent enhancing tumor from post-treatment changes. Restriction spectrum imaging (RSI), an advanced multishell diffusion technique, has shown promise in distinguishing cellular tumor from edema and enhancing post-treatment changes including pseudoprogression and radiation necrosis. We developed a convolutional network (nnUNet) that incorporates traditional multimodal MRI with DSC perfusion and RSI to detect and quantify enhancing and nonenhancing cellular tumor in post-treatment glioblastoma patients. This cellular tumor segmentation network was evaluated in a cohort of 186 post-treatment GBM patients (55.5 +/- 13.7 years, 158 male) with 243 timepoints using 5-fold cross validation. Cellular tumor volumes were then used to predict overall survival (OS) and progression free survival (PFS) in 94 patients with imaging within 3 months of initial surgery. Accuracy for segmenting cellular tumor as measured by the Dice-similarity coefficient was 0.77 (IQR 0.55-0.87) and the AUC of the ROC curve for detecting residual/recurrent tumor compared to post-treatment changes was 0.86. The predicted cellular tumor volumes were an independent hazard over age, sex, GTR status and conventional enhancing residual tumor for both OS (p=0.0003) and PFS (p=0.0004). The tumor segmentation network was validated in an external dataset of 22 GBM patients (55 timepoints) with a median Dice of 0.73 and AUC of 0.91. Survival prediction was validated in a different external dataset of 70 post-treatment GBM patients, with cellular tumor volumes being an independent predictor of OS (p=0.03). In conclusion, deep learning models incorporating advanced imaging can accurately segment enhancing and nonenhancing cellular tumor, which can help distinguish recurrent/residual tumor from post-treatment changes and predict OS and PFS.