2779

Abstract

An atlas-based automatic segmentation algorithm (ABAS) of brain OAR on MR images was developed by INRIA (National Institute Of Research in Applied Informatics) and integrated within ISOgray® (DOSIsoft software). We present its evaluation in clinical context. The use of the ABAS for CRT requires a CT/MR image registration. This registration process was first assessed. Then, we analyzed preliminary ABAS results in a retrospective study on 22 adult patients referred for CRT of their primary brain tumours. The block matching based registration software developed by INRIA allows for rigid transformation. Eight series of CT images of the Liqui-PhilTM Head Phantom (The Phantom Laboratory) rotated and/or tilted have been registered with MR images. The accuracy of the registration process has been evaluated by comparing the coordinates of 4 external fiducial markers on each pair of CT/MR registered images. Secondly, a CT/MR image registration has been performed and ABAS used for 22 patients treated in 2005: 11 high grade gliomas, 7 unfavourable low grade gliomas, 4 meningiomas (either resected or biopsied). Two expert clinicians used a “semi-quantitative” index from 0 to 5 to evaluate in terms of shape and position the automatic segmentation results for 8 OAR: eyes, optic nerves, optic chiasm, pituitary gland, cerebellum, brain stem. The image registration on phantom was very satisfactory for all the combinations: the maximum deviations in x, y and z directions were inferior to the image voxel size. The ABAS was fast and robust: 14 minutes were necessary to perform the segmentation of the 8 OAR for each of the 22 patients. Results were excellent for the eyes, the cerebellum and the brain stem: the index mean values and standard deviations were equal to 4.3 (± 0,24), 4.5 (± 0,48) and 4.8 (± 0,33) respectively. For small structures as pituitary gland, optic chiasm and optic nerves, the index mean values and standard deviations were equal to 3.25 (± 1,28), 3.75 (± 1,36) and 3.65 (± 1,16) respectively: due to the 3mm-slice thickness, these structures were often visible on a single slice. Through the phantom study, the registration software appeared to be very reliable. Preliminary results obtained with the ABAS were very encouraging: a prospective quantitative study of the ABAS is currently under achievement.