Epco-30. a radiomics based regression model to predict glioblastoma cell motility

Abstract PURPOSE Radiomics has shown considerable success as a diagnostic tool in the classification of tumor grade, stage, and prognosis. The purpose of this work is to examine the potential for radiomics techniques to accurately predict the motility properties of glioblastoma cells. METHODS Tissues specimens were obtained from a total of 31 patients undergoing surgical resection of glioblastoma. Mean tumor cell motility was calculated from time-lapse videos using previously described methods (Bangasser et al. Nat Comm 2017). Preoperative, post-contrast T1-weighted MR images with 1 mm voxels were obtained from the medical records for each patient. Manual segmentation was used to define the border of the enhancing tumor and 108 radiomics features were extracted from the normalized image volumes using the PyRadiomics software package. Imaging features that correlated strongly with cellular motility were selected in a stepwise manner by p-value. The four features most strongly correlated with cellular motility were included in a regression model using the adaptive lasso technique with leave-one-out cross validation (LOOCV). RESULTS Two first order and two gray-level run length matrix features were selected for the model. The R-squared value for the predictive model was 0.60 with p-values for each individual parameter estimate less than 0.0001. An analysis of the residual values of the predicted model did not show any evidence of bias in the estimate. The average root mean squared error between the predicted and actual motility from the LOOCV for the model was 0.75. CONCLUSION The results of this work suggest that it is possible for a quantitative image feature-based model to predict the cellular motility of glioblastomas. Further work will prospectively test the model and explore the role of cellular motility in clinical outcomes such as time to recurrence and patterns of failure.