Author: Beth M. Beadle, Adrian Celaya, Laurence Edward Court, David Fuentes, Anna Lee, Tze Yee Lim, Dragan Mirkovic, Amy Moreno, Raymond Mumme, Tucker J. Netherton, Callistus M. Nguyen, Jaganathan A Parameshwaran, Jack Phan, Carlos Sjogreen, Sara L. Thrower, Congjun Wang, He C. Wang, Xin Wang 👨🔬
Affiliation: Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center; The University of Texas MD Anderson Cancer Center UTHealth Houston Graduate School of Biomedical Sciences, Department of Radiation Oncology, Stanford University, The University of Texas MD Anderson Cancer Center, MD Anderson Cancer Center, MD Anderson, Rice University, Department of Radiation Physics, The University of Texas MD Anderson Cancer Center 🌍
Purpose: Quality assurance of target volumes from radiotherapy clinical trials is a labor and resource intensive task. The purpose of this work is to quantify the accuracy of a tool that automatically identifies the head and neck lymph node levels that are contained within manually contoured clinical target volumes (CTVs).
Methods: This tool is composed of a lymph node contouring model and an overlap assessment module. To develop ground truth data for the segmentation model, a radiologist contoured 28 lymph node levels on 128 CT scans of patients who previously received definitive radiotherapy per consensus guidelines. This dataset was used to train an 3D-full resolution nnU-Net model. Dice Similarity Coefficient (DSC) and mean surface distance (MSD) were calculated to assess model performance.
To curate ground truth data for the overlap assessment module, lymph node level contours were predicted on a separate cohort of 61 patient CTs from 4 different institutions. Then, a team of medical physicists manually recorded which of the 28 nodal levels comprised each elective CTV. Finally, Youden’s index was used to predict optimal DSC thresholds to identify lymph node levels which comprise elective CTVs. Five-fold cross validation accuracy, AUC, sensitivity, and specificity were calculated.
Results: The nnU-Net model could accurately contour all lymph node level contours with average DSC of 0.87+/-0.08 and average MSD [mm] of 0.64+/-0.41. Lymph node levels were accurately identified within elective CTVs of the multi-institutional cohort and demonstrated AUC of 0.94+/-0.06, accuracy of 0.87+/-0.09, sensitivity of 0.82+/-0.30, and specificity of 0.88+/-0.10 for patients in which surgical status, presence of intravenous contrast, and imaging protocol varied.
Conclusion: A highly accurate and comprehensive lymph node level contouring and identification tool was developed. The performance upon multi-institutional data indicated that this tool is robust and could be applied to CT scans from different clinical practices.