Does atlas-based autosegmentation of neck levels require subsequent manual contour editing to avoid risk of severe target underdosage? A dosimetric analysis
Background and purpose: To investigate the dosimetric impact of not editing auto-contours of the elective neck and organs at risk (OAR), generated with atlas-based autosegmentation (ABAS) (Elekta software) for head and neck cancer patients. Materials and methods: For nine patients ABAS auto-contours and auto-contours edited by two observers were available. Based on the non-edited auto-contours clinically acceptable IMRT plans were constructed (designated 'ABAS plans'). These plans were then evaluated for the two edited structure sets, by quantifying the percentage of the neck-PTV receiving more than 95% of the prescribed dose (V95) and the near-minimum dose (D99) in the neck PTV. Dice coefficients and mean contour distances were calculated to quantify the similarity of ABAS auto-contours with the structure sets edited by observer 1 and observer 2. To study the dosimetric importance of editing OAR auto-contours a new IMRT plan was generated for each patient-observer combination, based on the observer's edited CTV and the non-edited salivary gland auto-contours. For each plan mean doses for the non-edited glands were compared with doses for the same glands edited by the observer. Results: For both observers, edited neck CTVs were larger than ABAS auto-contours (p ≤ 0.04), by a mean of 8.7%. When evaluating ABAS plans on the PTVs of the edited structure sets, V95reduced by 7.2% ± 5.4% (1 SD) (p < 0.03). The mean reduction in D99was 14.2 Gy (range 1-54 Gy). Even for Dice coefficients >0.8 and mean contour distances <1 mm, reductions in D99up to 11 Gy were observed. For treatment plans based on observer PTVs and non-edited auto-contoured salivary glands, the mean doses in the edited glands differed by only -0.6 Gy ± 1.0 Gy (p = 0.06). Conclusions: Editing of auto-contoured neck CTVs generated by ABAS is required to avoid large underdosages in target volumes. Often used similarity measures for evaluation of auto-contouring algorithms, such as dice coefficients, do not predict well for expected PTV underdose. Editing of salivary glands is less important as mean doses achieved for non-edited glands predict well for edited structures.