Encrypted login | home

Program Information

PET-Based Radiomics to Predict Outcomes Following Definitive Chemoradiotherapy for Oropharyngeal Cancer

J Oh

J Oh1*, M Folkert2 , J Setton1 , A Apte1 , M Grkovski1 , R Young1 , H Schoder1 , W Thorstad3 , N Lee1 , J Deasy1 , (1) Memorial Sloan Kettering Cancer Center, New York, NY, (2) The University of Texas Southwestern Medical Center, Dallas, TX, (3) Washington University School of Medicine, St. Louis, MO


TH-AB-201-3 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 201

Purpose: We investigate whether PET-based radiomics features are correlated with outcomes in head-and-neck squamous cell cancer (HNSCC) patients with stage III-IV oropharyngeal cancer (OC) following definitive chemoradiotherapy.

Methods: There were 174 evaluable patients treated at our institution from 12/2002 to 3/2009. Using the CERR software, 24 shape, texture, and histogram-metric based features were extracted from pre-treatment FDG-PET scans. For all-cause mortality (ACM), local failure (LF), and distant metastases (DM) endpoints, the most frequently chosen logistic regression model during a leave-one-out cross validation in a manner of forward feature selection was determined as the best-fitting model. For an unbiased model estimate, the modeling process was tested using 5-fold cross validation. On the final predictive models, external validation tests were performed using an independent cohort, consisting of 65 patients with stage III-IV OC treated at another institution.

Results: Median followup time was 55 months (range 6-112 months). Among 174 patients, 48(27.6%) patients died, and 12(6.9%) and 33(19.0%) patients had LF and DM, respectively. For logistic regression models, ACM was correlated with kurtosis and metabolic tumor volume (MTV); LF was correlated with coherence and MTV; DM was correlated with solidity, kurtosis, and MTV. The 5-fold cross validation resulted in AUC=0.67 (p=0.024), 0.76 (p=0.039), and 0.66 (p=0.035) for ACM, LF, and DM, respectively. In external validation tests, significant predictive power was retained in LF with AUC=0.70 (p=0.024), and borderline significance was found for ACM and DM with AUC=0.60 (p=0.092) and 0.65 (p=0.062), respectively. To investigate the significance of MTV, patients were split into two groups by median MTV, and a Kaplan-Meier test was performed. Statistically significant differences were found for all three endpoints with p=0.034, 0.025, and 0.026 for ACM, LF, and DM, respectively.

Conclusion: We showed that predictive models designed using PET-based radiomics features can predict outcomes with a good level of performance.

Contact Email: