報告題目：Brain MRI: From Bench to Bedside
Our mission is to optimize the care of cancer patients by using imaging as a biomarker to accelerate drug development and promote successful and timely translation of cancer discovery into clinical practice. In this context, we have established the Center for Biomedical Imaging in Oncology with bidirectional exchanges between the mouse hospital at the Lurie Family Imaging Center, and the Clinical Imaging Research laboratory at the Cancer Center. Multiple co-clinical trials have been successfully carried out to facilitate translational cancer imaging for drug development and precision medicine.
Patient survival in high grade gliomas (HGG) remains poor. HGG is the largest and most genetically and phenotypically heterogeneous category of primary brain tumor. Numerous novel chemical, targeted molecular and immunoactive therapies in trial produce promising responses in a small disparate subsets of patients but which patient will respond to which therapy remains unpredictable. Reliable imaging biomarkers for prediction and early detection of treatment response and survival are a critical need in neuro-oncology. Methods: Brain tumor MRI 'deep features' extracted by a convolutional neural network (CNN) transfer learning were combined with features derived from an explicitly-designed radiomics model to search for MRI markers predictive of overall survival (OS) in HGG patients. Brain MRI of 128 HGG patients from the Cancer Genome Atlas (TCGA) were used as the discovery data set and the brain MRI of 52 local HGG patients as the validation data set. We extracted 8192 deep features and 348 radiomics features from each cohort and used feature selection and elastic net-Cox model to distinguish longer from shorter-term survivors. Results: The combined feaure framework was predictive of OS in both the discovery set (log-rank test p value=0.0227) and validation cohort (log-rank test p value=0.0148). Conclusion: Combined deep and radiomics features can predict OS in HGG patients and merit further stufy for reproducible prediction of treatment response.