Xintao Hu1,2, Geoffrey Young3, Stephen Wong1,4, Kelvin Wong1,4
1Department of Radiology, The Methodist Hospital Research Institute, Houston, TX, USA; 2Department of Automation, Northwestern Polytechnical University, Xi'an, ShaanXi, China; 3Department of Radiology, Brigham and Women's Hospital, Boston, MA, USA; 4Department of Radiology, Weill Medical College of Cornell University, New York, NY, USA
We proposed a method using cost-sensitive classifier to automatically identify radiation necrosis tissue on a voxel-by-voxel basis in resected GBM patients undergoing radiotherapy. The 8-dimentional feature vector was constructed from multiple MRIs. Classifier was modeled using One-class Support Vector Machine. Two parameters in training were optimized with criteria of area under ROC (AUROC). Threshold T for generating discrete classifier was determined according to unequal misclassification cost which could be personally sensitive. Discrimination of each feature was also measured using AUROC. The method was validated in a small cohort of resected GBM patients with confirmed non-progressing disease.