Room: AAPM ePoster Library
Purpose: Radiomics is a powerful quantitative imaging tool which has shown the ability to correctly classify tumor histology from routine clinical images. This work examines the potential for radiomic techniques to non-invasively and accurately predict the cellular motility properties of glioblastoma.
Methods: Preoperative T1-weighted MR images were acquired from 23 patients with glioblastoma whose tumors were resected. Glioblastomas were segmented manually and 137 radiomics features were extracted from each MRI volume using the PyRadiomics software package. Cell motility data was calculated from the analysis of time-lapse movies of patient tumor cells. The movies consisted of images taken at 15 minute intervals acquired over 10 hours. Adaptive-lasso regression was used to select features and estimate parameters where the radiomics features served as input and the cell motility value served as the target for the model. Given the number of samples in our dataset, the number of features in the model was limited to two. Leave-one-out cross validation (LOOCV) was used to validate the model.
Results: A statistically significant correlation was seen between a number of radiomics features and cell motility. The adaptive-lasso method selected the two top performing features from the radiomics dataset for use in the model: the grey-length-correlation-matrix maximum-probability and the first order 10th percentile features. The R-squared value for the model was 0.72. The p-value for the parameter estimates were both less than 0.0001. The average root mean squared error from the LOOCV for the model was 0.71.
Conclusion: This work suggests that radiomics features derived from medical imaging volumes contain information that characterizes tumors on a cellular level. Further work will test the robustness of our model in a prospective setting and will also evaluate whether knowledge of the cellular motility features in turn predicts modes of failure after treatment.