SU-C-BRA-05: Delineating High-Dose Clinical Target Volumes for Head and Neck Tumors Using Machine Learning Algorithms.

C. Cardenas,A. Wong,A. Mohamed,J. Yang,L. Court,A. Rao,C. Fuller,M. Aristophanous

DOI: https://doi.org/10.1118/1.4955566

2016-06-01

Abstract: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 3×3×3cm3 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. Varian Medical Systems grant.