Encrypted login | home

Program Information

Large-Scale Radiomics as Parsimonious Predictors Based On Imaging Eigen-Functions

no image available
S Tzeng

S Tzeng1*, S Harmon2 , T Perk2 , J Zhu2 , S Chen3 , R Jeraj2 , (1) China Medical University, Taichung, Taiwan, (2) University of Wisconsin, Madison, WI, (3) The 1st Hospital of China Medical University, Shenyang, Liaoning


MO-RAM-GePD-J(A)-2 (Monday, July 31, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Lounge - A

Purpose: Most existing texture features in radiomics focus on local spatial behaviors or ignore spatial structures. This study develops a novel method to extract data-driven features accounting for more global spatial characteristics of imaging, which may lead to a better understanding of the spatial patterns they capture.

Methods: The model utilizes image intensities within each region of interest (ROI) from patients. Simulations mimicking phantom scans were conducted with varying image complexity, resolution, and sample size. Then variation in the number of extracted eigen-components was investigated. The novel features were applied to a study of solitary pulmonary nodules from 85 patients’ [F-18]FDG dual time point PET images (DTPI). Each patient was scanned twice for first and delayed imaging. Patients were classified as benign or malignant, subsequently confirmed by pathological results or follow-ups (22 benign and 63 malignant lesions). Classification performances of support vector machines trained using two different feature sets were compared: (1) proposed features, (2) 120 common imaging features.

Results: The method decomposes image intensities into linear combinations of eigen-functions, and the weights for linear combinations are then used as data-driven features. An expectation-maximization algorithm iteratively learns both eigen-functions and the weights, while the optimal number of functions is determined by Akaike information criterion. We found the novel data-driven features often describing large spatial characteristics. The number of data-driven features increases with image resolution and complexity but is insensitive to sample size, which is a desirable property in real application. Only 4 proposed data-driven features were found in the DTPI data. Out-of-sample AUCs in cross validation for the two feature sets were (1) 0.820 and (2) 0.838.

Conclusion: The few large-scale features based on eigen-functions performed almost as well as a large number of texture features, implying the parsimonious explanatory ability of the data-driven method.

Funding Support, Disclosures, and Conflict of Interest: Global Networking Talent 3.0 Plan

Contact Email: