Purpose: To evaluate the geometrical and associated dosimetric accuracy of an Atlas – Based Autosegmentation software (ABAS 2.01, Elekta) on organs-at-risks (OARs) for Tomotherapy planning of nasopharyngeal carcinoma (NPC) patients.
Methods: 20 NPC patients who received Tomotherapy treatment were recruited retrospectively. For each patient, the ABAS software was used to generate auto-segmented contours on 20 OARs. Geometric metrics including dice similarity coefficient (DSC), Hausdorff distance 95th-percentile (HD95) and mean surface distance (MSD) were used to evaluate the geometric agreement between the auto-segmented contours and contours delineated manually by experienced medical staff. Tomotherapy based on auto-segmented contours with DSC>0.7 and the original PTVs were planned for each patient. The dose distributions were evaluated on both auto- and manually-segmented contours, and the dosimetric parameters were compared. Correlations between absolute dose differences and geometric metrics were evaluated using Spearman’s correlation analysis.
Results: The software generated contours with acceptable geometric accuracy for 11 OARs (DSC>0.7), including mandible, eye, tongue, temporal lobe, brainstem, larynx, esophagus and spinal cord. However, temporal lobe and esophagus showed large boundary mismatch (HD95>5mm). Contouring performance was worst for brachial plexus (DSC=0.44, HD95=8.6mm). Dosimetrically, the mean doses of mid-ear, supraglottic larynx, larynx and esophagus showed no significant differences between auto- and manually-segmented contours. Significant correlations were observed between mean dose differences of external auditory canal, mid-ear, larynx and their corresponding geometric metrics.
Conclusion: The ABAS software demonstrated acceptable performance in some of the NPC OARs, but showed limited geometric accuracy for organs with poor image contrast and small volume. No significant dosimetric differences were shown between auto- and manually-segmented contours for some organs, which may imply that further manual adjustment to these auto-segmented contours may not be necessary. Finally, for certain organs, the geometric metrics may be a good predictor for dosimetric accuracy of the autosegmentation owing to their strong correlation.