Program Information
Deep Learning On Clinically-Clustered Patients Improves Auto-Delineation of Oropharyngeal High-Risk Clinical Target Volumes
C Cardenas*, R McCarroll , L Court , B Elgohari , H Elhalawani , C Fuller , M Jomaa , M Meheissen , A Mohamed , A Rao , B Williams , A Wong , J Yang , M Aristophanous , UT MD Anderson Cancer Center, Houston, TX
Presentations
MO-L-GePD-J(B)-3 (Monday, July 31, 2017) 1:15 PM - 1:45 PM Room: Joint Imaging-Therapy ePoster Lounge - B
Purpose: Radiation oncologists consider clinical information when manually contouring clinical target volumes (CTVs). In this study, we investigate whether the inclusion of patient clinical information improves the auto-delineation of high-risk CTVs.
Methods: Simulation CT images and clinical information from forty-one oropharyngeal patients were collected. Clinical features collected included tumor site, T/N stage, pathological and tumor-specific information. Patients were separated by their clinical features into clusters using Ward’s hierarchical clustering. Target volumes and normal tissues in the head-and-neck region were manually contoured by five radiation oncologists. Voxel-based classification was performed via deep neural networks (DNN) using spherical coordinates to the nearest voxel from each contoured anatomical structures as predictors to determine if the voxel was part of the CTV. Training was performed using patients within each clinical cluster. For comparison, 10 patients were randomly chosen and prediction models were trained using the same methodology as for the clustered patients. Leave-one-out cross-validation was used to train the models for each cluster and to predict on the non-clinically clustered patients. Volume metrics were calculated to compare auto-delineated CTVs to physician manual contours.
Results: Patients were clustered in 6 groups with a median of 6 patients per clinical cluster (range: 4-11). Overall, clinically-clustered DNN predicted volumes showed higher agreement to the manual contours (p = 0.026, paired-Wilcoxon rank-sum test). The median Dice similarity coefficients (DSC) were 0.827 (range: 0.752-0.894) and 0.817 (range: 0.472-0.888) for the clinically-clustered and non-clustered predicted volumes, respectively. The average improvement in DSC using the clinically-clustered volumes was 0.022 (range: -0.056 to 0.378). On average, DNN training took 15 hours per patient while predictions were completed in 0.001 seconds using an NVIDIA K40 GPU.
Conclusion: Our data shows that clinically clustering patients prior to deep learning improves prediction and shows promise in predicting physician behavior when auto-delineating high-risk clinical target volumes.
Funding Support, Disclosures, and Conflict of Interest: This work was funded by a Varian MRA grant. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Tesla K40 GPU and the Texas Advanced Computing Center (TACC) at The University of Texas at Austin for providing computational resources.
Contact Email: