Program Information
Delineating High-Dose Clinical Target Volumes for Head and Neck Tumors Using Machine Learning Algorithms
C Cardenas1,2*, A Wong3,4 , A Mohamed3 , J Yang1 , L Court1 , A Rao5 , C Fuller3 , M Aristophanous1 , (1) Department of Radiation Physics, The University of Texas M.D. Anderson Cancer Center, Houston, TX, (2) The University of Texas Graduate School of Biomedical Sciences, Houston, TX (3) Department of Radiation Oncology, The University of Texas M.D. Anderson Cancer Center, Houston, TX, (4) School of Medicine, The University of Texas Health Sciences Center at San Antonio, San Antonio, TX, (5) Department of Bioinformatics and Computational Biology, The University of Texas M.D. Anderson Cancer Center, Houston, TX
Presentations
SU-C-BRA-5 (Sunday, July 31, 2016) 1:00 PM - 1:55 PM Room: Ballroom A
Purpose: To develop and test population-based machine learning algorithms for delineating high-dose clinical target volumes (CTVs) in H&N tumors. Automating and standardizing the contouring of CTVs can reduce both physician contouring time and inter-physician variability, which is one of the largest sources of uncertainty in H&N radiotherapy.
Methods: Twenty-five node-negative patients treated with definitive radiotherapy were selected (6 right base of tongue, 11 left and 9 right tonsil). All patients had GTV and CTVs manually contoured by an experienced radiation oncologist prior to treatment. This contouring process, which is driven by anatomical, pathological, and patient specific information, typically results in non-uniform margin expansions about the GTV. Therefore, we tested two methods to delineate high-dose CTV given a manually-contoured GTV: (1) regression-support vector machines(SVM) and (2) classification-SVM. These models were trained and tested on each patient group using leave-one-out cross-validation. The volume difference(VD) and Dice similarity coefficient(DSC) between the manual and auto-contoured CTV were calculated to evaluate the results.
Distances from GTV-to-CTV were computed about each patient’s GTV and these distances, in addition to distances from GTV to surrounding anatomy in the expansion direction, were utilized in the regression-SVM method. The classification-SVM method used categorical voxel-information (GTV, selected anatomical structures, else) from a 3x3x3cm3 ROI centered about the voxel to classify voxels as CTV.
Results: Volumes for the auto-contoured CTVs ranged from 17.1 to 149.1cc and 17.4 to 151.9cc; the average(range) VD between manual and auto-contoured CTV were 0.93 (0.48-1.59) and 1.16(0.48-1.97); while average(range) DSC values were 0.75(0.59-0.88) and 0.74(0.59-0.81) for the regression-SVM and classification-SVM methods, respectively.
Conclusion: We developed two novel machine learning methods to delineate high-dose CTV for H&N patients. Both methods showed promising results that hint to a solution to the standardization of the contouring process of clinical target volumes.
Funding Support, Disclosures, and Conflict of Interest: Varian Medical Systems grant
Contact Email: