Program Information
Quantifying Daily Changes in Cone-Beam CT Radiomics to Predict Radiation-Induced Long-Term Xerostomia
B Rosen1*, K Brock2 , C Lockhart1 , J Kamp1 , A Eisbruch1 , R Ten Haken1 , I El Naqa1 , (1) University of Michigan , Ann Arbor, MI, (2) UT MD Anderson Cancer Center, Houston, TX
Presentations
TH-AB-201-7 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 201
Purpose: To assess whether daily CBCT image features can improve long-term xerostomia prediction during head and neck radiotherapy.
Methods: Deformable registrations of planning CTs to daily CBCTs were performed using Varian SmartAdapt for 162 head and neck cancer patients who received a primary radiotherapy course (70 Gy/35 fractions). Deformed parotid gland structures were propagated to subsequent CBCT images, and intensity histogram, shape, and texture features were extracted using in-house software. RTOG xerostomia was scored at regular follow-ups. Baseline gland volume, mean dose, and changes in radiomic features over the treatment course were used in prediction models for one-year Grade 2 or higher xerostomia. A total of 106 features were extracted for each patient fraction, and sequential forward selection using bootstrap replication was used for parameter reduction. Further, any features predictive of the clinical endpoint derived from the last 15 fractions were simulated using linear least-squares regression of the first 20 fractions to evaluate clinical utility.
Results: 92 patients (57%) had the full parotid gland within the CBCT field-of-view over the entire treatment course, of which 46 (50%) had sufficient follow-up and eight (17%) had Grade 2 xerostomia. Although mean dose was correlated with acute volume change (R²=0.154), neither mean dose nor volume change was predictive of long-term xerostomia (AUC=0.618 (95%CI: 0.387-0.825), p=0.521). Conversely, measured and simulated end-treatment NGTDM-Complexity texture was strongly predictive (AUC=0.803 (CI: 0.659-0.922), p<0.004 and AUC=0.740 (CI: 0.586-0.880), p<0.02, respectively). When mid-treatment changes in GLRLM-RLV were added, performance of the texture model improved (AUC=0.852 (CI: 0.724-0.952), p<0.005). Further, visual inspection of parotid images over time revealed noticeable variations in duct junctions, which may provide a physical explanation of the derived radiomics signature.
Conclusion: CBCT radiomics may potentially identify the onset of irreversible radiation-induced damage to the parotids and help support adaptive radiotherapy in head and neck cancer.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by NIH grant P01CA059827.
Contact Email: