Purpose
In online cone beam computed tomography (CBCT)-based adaptive radiation therapy (ART), nodal recontouring ensures sufficient nodal coverage by accounting for anatomic change but is uniquely challenging due to small target size and time pressure. This study evaluates the accuracy of rigid propagation versus artificial intelligence-guided deformation (AID) for nodal autosegmentation via comparison to nodal recontours delineated with unlimited time (ie, benchmark contours).
Methods and Materials
We analyzed 25 nodal structures from 16 patients receiving pelvic online CBCT-based ART with nodal boost. Nodal structure sampling was informed by an initial power analysis. For each structure, we obtained rigidly propagated and AID-generated contours in addition to 2 benchmark contours and the clinical contour used in adapted plan generation. We calculated dice similarity coefficient (DSC), false-positive dice, false-negative dice, and 95% Hausdorff distance (HD95) between clinical, propagated, and AID contours against benchmark pairs and DSC and HD95 between benchmark pairs. The failure rate of nonbenchmark contours relative to benchmark pairs was calculated as the proportion of contours with an HD95 > 5 mm. We calculated the normalized D100, normalized D95, V100, and V95 of the adapted plan dose over all contours. Clinical tumor volume contours were used for all comparisons.
Results
Median DSC versus benchmark contours were 0.68 for rigidly propagated and 0.58 for AID contours. A significant difference in false-negative dice (P = .01, Cohen’s d 0.806) was identified in benchmark-to-propagated versus benchmark-to-AID comparison. The failure rate of rigidly propagated, AID, and clinical contours was 20%, 48%, and 28% respectively. The median DSC between benchmark contours was 0.75. No significant differences across dose metrics were identified between contour types.
Conclusions
Rigid propagation is superior to AID for initial contour generation in pelvic CBCT-based ART. Increased contouring time and image quality may improve contour quality and reduce interobserver variability but may be limited by the influence of individual contouring preferences.
扫码关注我们
求助内容:
应助结果提醒方式:
