Automatic segmentation for adaptive planning in nasopharyngeal carcinoma IMRT: Time, geometrical, and dosimetric...

The aim of this study was to quantify the geometrical differences between manual contours and autocontours, the dosimetric impacts, and the time gain of using autosegmentation in adaptive nasopharyngeal carcinoma (NPC) intensity-modulated radiotherapy (IMRT) for a commercial system. A total of 20 consecutive Stages I to III NPC patients who had undergone adaptive radiation therapy (ART) re planning for IMRT treatment were retrospectively studied. Manually delineated organs at risks (OARs) on the replanning computed tomography (CT) were compared with the autocontours generated by VelocityAI using deformable registration from the original planning CT. Dice similarity coefficients and distance-to-agreements (DTAs) were used to quantify their geometric differences. IMRT test plans were generated with the assistance of RapidPlan based on the autocontours of OARs and manually segmented target volumes. The dose distributions were applied on the manually delineated OARs, their dose volume histograms and dose constraints compliances were analyzed. Times spent on target, OAR contouring, and IMRT replanning were recorded, and the total time of replanning using manual contouring and autocontouring were compared. The averaged mean DTA of all structures included in the study were less than 2 mm, and 90% of the patients fulfilled the mean distance agreement tolerance recommended by AAPM 132.¹ The averaged maximum DTA for brainstem, cord, optic chiasm, and optic nerves were all less than 4 mm, whereas temporal lobes and parotids have larger average maximum DTA of 4.7 mm and 6.8 mm, respectively. It was found that large contour discrepancies in temporal lobes and parotids were often associated with large magnitude of deformation (warp distance) in image registrations. The resultant maximum dose of manually segmented brainstem, cord, and temporal lobe and the median dose of manually segmented parotids were found to be statistically higher than those to their autocontoured counter parts in test plans. Dose constraints of the manually segmented OARs in test plans were only met in 15% of the cases. The average time of manual contouring and autocontouring were 108 and 10 minutes, respectively (p < 0.001). More than 30% of the total replanning time would be spent in manual OAR contouring. Manual OAR delineation takes up a significant portion of time spent in ART replanning and OAR autocontouring could considerably enhance ART workflow efficiency. Geometrical discrepancies between auto- and manual contours in head and neck OARs were comparable to typical interobserver variation suggested in various literatures; however, some of the corresponding dosimetric differences were substantial, making it essential to carefully review the autocontours.