Click here to

Session: Data Science, Radiomics, and Computing [Return to Session]

JACK KROHMER EARLY-CAREER INVESTIGATOR COMPETITION WINNER: Multi-Group Multi-Block Data Integration for Harmonizing 18F-FDG-PET/CT Radiomics Associated with Circulating Tumor Cells and Predicting Recurrence-Free Survival Across Independent Lung Cancer Radiotherapy Studies

S Lee*, G Kao, S Feigenberg, Y Fan, Y Xiao, University of Pennsylvania, Philadelphia, PA


SU-E-TRACK 6-3 (Sunday, 7/25/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: To combine independent studies with a multi-group multi-block (MGMB) integrative method for harmonizing ¹⁸F-FDG-PET/CT radiomics associated with circulating tumor cells (CTCs), and evaluate the impact on predicting recurrence-free survival (RFS) after radiotherapy in non-small cell lung cancer (NSCLC).

Methods: The present study investigated two independent studies, consisting of 56/50 patients with early-stage/locally-advanced NSCLC (ES/LA-NSCLC) treated with stereotactic body radiotherapy (SBRT)/definitive concurrent chemoradiotherapy (CCRT), with pretreatment ¹⁸F-FDG-PET/CT imaging and CTC assessment via a telomerase-based assay. Each study included 1,548/1,562 PET/CT radiomic features extracted from gross tumor volume plus 5 baseline clinical parameters (CPs). MGMB data integration was performed to integrate four (CT/PET/CP/CTC)-block feature sets and combine both studies (ES-NSCLC SBRT and LA-NSCLC CCRT) by using sparse MGMB generalized partial least squares (PLS) regression with unsupervised feature selection (UFS). Algorithm for RFS data stacking was implemented for RFS analysis as a classification problem. Block PLS-discriminant analysis (PLS-DA) was performed to integrate the four-block feature sets and predict locoregional/distant RFS after radiotherapy, where one study group served as a training set while the other as a testing set. The same analysis was conducted via single-group multi-block (SGMB) data integration that naively concatenated both studies while integrating the four-block feature sets by using sparse MB generalized PLS regression with UFS. Predictive performance of the block PLS-DA on the testing set for locoregional/distant RFS was compared using Harrell's c-index between MGMB and SGMB integrative approaches.

Results: For predicting locoregional/distant RFS, the MGMB vs. SGMB integration obtained c-indices of 0.653/0.609 vs. 0.547/0.530 for the pre-SBRT ES-NSCLC study and c-indices of 0.736/0.758 vs. 0.730/0.882 for the pre-CCRT LA-NSCLC study, respectively.

Conclusion: The MGMB integration of ¹⁸F-FDG-PET/CT radiomics and CTC improved predictive performance through exploring potential correlation between studies, indicating that the MGMB integration may reduce prediction bias for RFS when combining independent NSCLC radiotherapy studies.

Funding Support, Disclosures, and Conflict of Interest: This project was supported by NCI grants, U24CA180803(IROC) and U10CA180868(NRG). Dr. Kao is a co-founder and has equity in Liquid Biotech USA, Inc., a University of Pennsylvania PCI-developed company through the UPStart program.



    Image Analysis, Lung, Radiation Therapy


    IM/TH- Image Analysis Skills (broad expertise across imaging modalities): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

    Contact Email