Encrypted login | home

Program Information

BEST IN PHYSICS (JOINT IMAGING-THERAPY) - Modeling Pathologic Response of Locally Advanced Esophageal Cancer to Chemoradiotherapy Using Spatial-Temporal FDG-PET Features, Clinical Parameters and Demographics

H Zhang

H Zhang1*, S Tan2, W Chen1, S Kligerman1, G Kim1, W DSouza1, M Suntharalingam1, W Lu1, (1) University of Maryland School of Medicine, BALTIMORE, MD, (2) Huazhong University of Science and Technology

WE-C-BRA-1 Wednesday 10:30:00 AM - 12:30:00 PM Room: Ballroom A

Purpose: To develop tumor response models using quantitative spatial-temporal FDG-PET features to accurately and precisely predict pathologic response to concurrent chemoradiotherapy (CRT) for locally advanced esophageal cancer patients.

Methods: Data from 20 patients who underwent tri-modality therapy (CRT plus surgery) were retrospectively evaluated. Patients underwent FDG-PET/CT scans before initiation of CRT and 4-6 weeks after completion of CRT but prior to surgery. Three types of image features were examined for their ability to predict pathologic tumor response: (1) conventional PET/CT response measures (SUVmax, SUVpeak, tumor diameter); (2) clinical parameters (TNM stage, histology, etc) and demographics (patient age, sex, etc.); (3) spatial-temporal PET features, which characterize the FDG uptake distribution, spatial variation (texture), geometric properties of a tumor and their temporal changes. Recursive feature selection was used to identify an optimal subset of features from the full feature set. Support vector machine (SVM) and logistic regression models were constructed for the prediction using the optimal subset as input. Cross-validation with randomized training and testing data were used to avoid over-fitting. Model accuracy was assessed via sensitivity, specificity and receiver operating characteristics (ROC) curve analysis. Model precision was assessed via confidence intervals of the mean.

Results: Highly accurate pathologic tumor response was obtained using SVM with an area under the ROC curve (AUC) of 0.95, (100% sensitivity and 90% specificity). The 95% confidence interval was [0.92, 0.98], suggesting high precision as well. The improvements using spatial-temporal PET features over only using conventional PET/CT response measures (AUC=0.82) or clinical parameters and demographics (AUC=0.77) were significant (p<0.0001).

Conclusions: By using quantitative PET/CT image features we were able to accurately and precisely predict pathologic tumor response in esophageal cancer patients undergoing concurrent CRT. This prediction has the potential to alter current clinical practice standards by identifying patients in whom surgery may be safely deferred.

Contact Email