P122. Multimodal probabilistic atlas of the human inner ear enables an automated and reliable segmentation of the endolymphatic space

Introduction Intravenous contrast agent enhanced magnet resonance imaging (ivMRI) of the endolymphatic space (ELS) of the inner ear permits the in vivo non-invasive verification of the endolymphatic hydrops (ELH). This might prove especially helpful in the at times complex clinical differentiation between Vestibular Migraine (VM) and Menieres Disease (MD) ( Nakada et al., 2014 ). However, current quantification methods of the ELS are cumbersome and experience-dependent ( Gurkov et al., 2015 ). The aim of this study was (i) the development of a multimodal probabilistic atlas of the bony structures of the human inner ear, (ii) the establishment of an atlas-based segmentation of the ELS to then (iii) enable an automated and reproducible volume-based quantification of the ELS. Methods Twenty-four patients with vestibular migraine, but without ELH (12 females, aged 20 −76 years, mean age: 51.5 ± 3.9 years) underwent a delayed MRI of the ELS 4 h after the intravenous injection of gadobutrol was performed in a 3 T scanner ( Nakada et al., 2014 ). Diagnostic work-up included neurologic & neuro-orthoptic assessment, VOG during caloric stimulation and head-impulse [HIT], audiometry. The preprocessing is performed by an affine and deformable fusion pipeline and the probabilistic multimodal atlas was created using a label-based approach ( Brandt and Dieterich, 2017 ). Results The performance of the atlas was evaluated using three factors: population, number of control points and second Gaussian kernel. The optimum value of the factors for the used algorithm was recognized and atlas based segmentation was conducted. The accuracy of the segmentation was assessed using a dice score (right and left side mean dice score 89% and 86%). Right and left inner ear probabilistic atlas resulted in a 4.7% discrepancy. Discussion This novel multimodal probabilistic atlas of the human inner ear enables an automated and reliable segmentation of the endolymphatic space using a deformable registration pipeline, in which each newly generated template includes more comprehensive and plenary information. Applications include automated 3D segmentation and quantification of the inner ear and its different fluid-filled spaces, as well as of the vestibulocochlear nerve.