Program Information
First Experience On Validation of Pathologic Response Prediction Models Utilizing PET/CT Radiomic Features Using Multi-Institutional Datasets
H Zhang1*, S Riyahi Alam2 , W Choi2 , D Ayala-Peacock3 , E McTyre4 , J Bourland4 , A Blackstock4 , W D'Souza1 , W Lu2 , (1) University of Maryland School of Medicine, Baltimore, MD, (2) Memorial Sloan Kettering Cancer Center, New York, NY, (3) Vanderbilt University Medical Center, Nashville, TN, (4) Wake Forest Baptist Medical Center, Winston Salem, NC
Presentations
SU-H1-GePD-J(B)-2 (Sunday, July 30, 2017) 3:00 PM - 3:30 PM Room: Joint Imaging-Therapy ePoster Lounge - B
Purpose: To validate tumor response prediction models that utilize FDG-PET/CT radiomic features for locally advanced esophageal cancer patients after chemo-radiotherapy (CRT) using multi-institutional datasets.
Methods: Datasets from two institutions including 41 patients (20 from institution A and 21 from institution B) who underwent tri-modality therapy (concurrent CRT plus surgery) were retrospectively evaluated. Patients underwent PET/CT scans before initiation of CRT and 4-6 weeks after completion of CRT but prior to surgery. Pathologic tumor response was used as ground truth to examine the predictability of radiomic features, clinical parameters and demographics. 79 and 80 radiomic features were extracted from (pre- and post-CRT) PET and CT, respectively. The feature temporal changes were calculated as well resulting to a total of 477 radiomic features. Feature selection and supervised machine learning were applied in 3 ways: (1) Using cross-validation for each institutional dataset individually; (2) Using cross-validation for the combined multi-institutional dataset and (3) using one institutional dataset for training and evaluating on another institutional dataset. Three ML models were tested, logistic regression (LR), support vector machine (SVM) and neural networks (NN). Model accuracy was assessed via receiver operating characteristics (ROC) curve analysis and confidence intervals of the mean.
Results: For two single institutional datasets, the most accurate tumor response predictions were obtained using NN with an area under the ROC curve (AUC) of 0.96 [0.93 0.99] and 0.90 [0.87 0.93], respectively. When combined as a multi-institutional dataset, SVM achieved the best performance, AUC = 0.98 [0.97 0.99]. The AUC was below 0.65 when using one institutional dataset for training and evaluating on another institutional dataset.
Conclusion: Our initial experience shows that high accuracy can be achieved when data from different institutions were combined. However, it is difficult to predict the response of one institution using another due to discrepancy between optimal feature sets.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the National Cancer Institute Grants R01CA172638.
Contact Email: