Hierarchical Clustering and Classification Improve the Prediction of GTV Response Throughout the Course of Head and Neck IMRT
A Yock1*, A Rao1, L Dong2, B Beadle1, L Zhang1, J Yang1, R Kudchadker1, L Court1, (1) UT MD Anderson Cancer Center, Houston, TX, (2) Scripps Proton Therapy Center, San Diego, CA.SU-E-J-211 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
To better predict the changing oropharyngeal tumor volume throughout the course of treatment for response assessment and potential adaptive intervention.
Nineteen patients with an aggregate of 35 GTVs (primary and nodal volumes) received daily CT-on-rails imaging for image-guided IMRT of the oropharynx. Daily images were registered to planning CTs using deformable image registration. Resulting vector fields were used to propagate contours from the planning CT to each daily CT. The time course of the changing volume of each GTV was smoothed by Gaussian process regression. Each GTV was selected as a test case, while the remaining 34 were used to generate class models using principal component analysis and hierarchical clustering. A time course of each class model was generated from the median principal component score of that class. For each test case, the accuracy of the optimal class model was compared with that of a static model assuming no change in volume from the planning CT, and with that of a linear model assuming GTV volume decreased at a constant rate. Subsets of test case daily imaging data were used to represent daily, every other day, weekly, and biweekly imaging schemes. For each scheme, the accuracy of classifying the test case as the optimal class model was evaluated at various treatment times.
The root mean square error of the optimal class model was less than that of the static and linear models by 10.1% and 9.7% of the initial GTV volume, respectively. This improved accuracy was statistically significant at the p<0.001 level. The accuracy of the classification was around 50% at the start of treatment and increased with the number of treated fractions. It did not vary considerably across imaging schemes.
Class solutions predict the GTV volume throughout treatment with accuracy greater than conventional clinical methods.
Funding Support, Disclosures, and Conflict of Interest: The presenting author acknowledges financial support from: 1) The University of Texas Graduate School of Biomedical Sciences at Houston; The University of Texas MD Anderson Cancer Center, Houston, TX 77030, 2) American Legion Auxiliary Fellowship, The University of Texas Graduate School of Biomedical Sciences at Houston, Houston, TX 77030.