Validation of an Automated Segmentation Method for Head and Neck Adaptive Radiotherapy using Cone-beam Computed Tomography (CBCT)

Abstract

An efficient and robust deformable registration method to propagate contours from the planning computed tomography (CT) to the daily online CBCT is needed to facilitate adaptive radiotherapy in clinical practice. The goal of this study is to quantitatively validate an automatic segmentation method based on 3D salient interest points on daily CBCT of the head and neck. Ten patients were identified who received radical intensity-modulated radiation therapy (IMRT) for head and neck cancer and presented with bulky neck nodes that were radiologically evident on CBCT. Ten observers (six radiation oncologists, three fellows and one specialized dosimetrist) volunteered for this study. For each patient, the observers manually delineated one nodal volume, the cervical spinal cord and one parotid on the planning CT images and on 3 CBCT images acquired in the first, middle and last week of radiotherapy. The observers evaluated the manually delineated structures on the planning CT and chose representative contours as consensus structures for each patient. These consensus structures were automatically propagated to each CBCT image. For each structure in each CBCT, a reference contour was created consisting of the sum of all observers' contours. Finite element modeling was used to quantify the mean displacement of surface elements from the reference contour to every contoured structure. The average of the mean displacements across all observers was compared with the automatic segmentation for all patients using a Wilcoxon rank test for each structure in each image. The auto-segmentation method took less than 2 minutes to propagate all consensus structures from the planning CT images to the 3 CBCT images. Qualitatively, all propagated nodal volumes and cords successfully followed the anatomic deformations as seen on the CBCT images; 5 out of 30 propagated parotids presented a small overlap into bone and/or beyond the patient outline. According to the Wilcoxon rank test for each CBCT, no statistically significant difference was observed between the auto-segmented and manual contours the nodal volumes, parotid and cord(α = 0.05). The results from this study demonstrated that the 3D salient interest point based automated segmentation is a reliable and efficient method to propagate large series of contours from a planning CT image to subsequent CBCT images with anatomic deformations. The automatically propagated contours were consistent with the observers' contours across all CBCT images. This provides a systematic framework to quantify the dosimetric effect of anatomic and geometric changes during a course of fractionated radiotherapy, to determine whether there is a need for replanning.