The Impact of the Number of Subjects for Atlas-Based Automatic Segmentation
J L Ducote*, V Sehgal, J Wong, M Al-Ghazi, University of California Irvine, Orange, CASU-E-J-102 Sunday 3:00:00 PM - 6:00:00 PM Room: Exhibit Hall
Purpose: To determine the impact of atlas size on the performance of atlas-based automatic segmentation (ABAS) in delineation of organs at risk for adaptive radiation therapy.
Methods: A total of 25 patients who had undergone intensity modulated radiation therapy for various head and neck cancers were retrospectively selected for inclusion in a library to be used for ABAS with the MIM VISTA software package (MIM Software, Cleveland OH). Treatment planning computed tomography (CT) scans and subsequent organ at risk (OAR) contours generated as part of the treatment planning process for these patients were added to the library. This library of 25 patients was then successively pruned to generate 5 atlases with 25, 20, 15, 10, and 5 patient subjects respectively. Atlas based segmentation was performed on 10 retrospectively selected treatment planning CT scans to automatically generate right and left parotid glands and brainstem contours. These planning CT scans belonged to a unique set of 10 patient subjects different from the ones used for generating the atlases. One physician (JW), who was blinded to the ABAS results, manually delineated gold-standard contours for the right and left parotid glands and brainstem. Dice similarity coefficients were calculated and analyzed as a function of atlas subject size.
Results: For the sites selected in this study, the performance of ABAS was relatively insensitive to atlas size. Furthermore, some patient subjects were repeatedly selected implying that the adoption of a single standard patient for ABAS may be of benefit.
Conclusions: Our preliminary results indicate that the performance of the atlas based segmentation module in MIM VISTA Version 5.2 for the organs studied here may be relatively insensitive to the atlas size.