Optimizing spatial normalization of multisubject inner ear MRI: comparison of different geometry-preserving co-registration approaches

Abstract

Spatial normalization of multisubject inner ear imaging data is challenging, due to both substantial intraindividual differences and the small size of the organ compared to other intracranial structures. Automatic whole brain co-registration to standard space can only roughly co-align the peripheral vestibular endorgan, and complemental manual registration is highly time-consuming. Here, we compared the accuracy of four geometry-maintaining co-registration methods (one semi-manual method and three automatic methods). High-resolution structural T2-MRI of 153 inner ears from patients and healthy participants were co-registered to an inner-ear atlas. The semi-manual method used a three-point landmark-based approach (3P), two automatic methods were based on unassisted standard algorithms (Advanced Normalization Tools (ANTs), Elastix (EL)), while the fourth automatic method utilized a volumetrically dilated, atlas-based mask (thick inner ear, TIE) for probabilistic inner ear masking. Registration accuracy was evaluated by neurotologists blinded to the respective registration paradigm, and the resulting median volumes were quantified using colocalization analyses. The mask-aided automatic approach showed the best ratings, followed by the semi-manual three-point landmark-based registration (mean ratings (lower: better) TIE 2.21 ± 1.15; 3P 2.58 ± 0.61; EL 3.42 ± 1.06; ANTs 3.49 ± 1.26). The semi-manual method had the lowest rate of insufficient registrations, followed by TIE (3P: 3.70%; TIE: 8.28%; EL: 22.66%; ANTs: 27.02%). TIE showed the highest colocalization metrics with the atlas. Only TIE and 3P allowed for sufficient semicircular canal visualization in method-wise average volumes. Overall, geometry-preserving spatial normalization of multisubject inner ear imaging data is possible and could allow groupwise examinations of the bony labyrinth or temporal bone morphology in the future.

Keywords: Atlas; Endolymphatic hydrops; Inner ear; Peripheral vestibular system; Spatial normalization; Vertigo.

Fig. 1

Visualization of the inner ear currently available in the MNI152, based on whole-brain coregistration from 152 healthy adults. While the resulting atlas has been used in countless neuroimaging studies as a standard coordinate system, the inner ear anatomy is not depicted in a sufficient resolution. (A) Overview of the MNI high resolution T2 atlas and zoom into the inner ear region. (B) 3D-rendering of the inner-ear anatomy available in the MNI152.

Fig. 2

Overview of the preprocessing, registration, and analysis steps for an individual dataset. After selecting both inner ears from the raw T2-SPACE, the left inner ear is mirrored along the sagittal plane to match the orientation of the T2-atlas. For every inner ear crop, two unaided automatic registrations (Elastix, ANTs), one semimanual registration (3P; using three landmarks: (1) apex cochlea, (2) lateral pSCC, (3) superior sSCC), and one mask-aided approach (TIE) were performed (target volume: atlas). While the landmark selection for the 3P-approach was done for every individual dataset, the atlas-derived dilated mask was only created once and used for all subsequent registrations. Registration accuracy was then rated by three experienced neurotologists blinded to the method used. Furthermore, statistical colocalization analysis on the masked registration was performed (not depicted). ANTs advanced normalization tools, 3P three point alignment, TIE thick inner ear mask coregistration.

Fig. 3

Average ratings of three independent raters, blinded to the method (1 = best possible grade, 6 = worst possible grade; lower average scores equal to better performance). On average, TIE showed the best ratings; however, in some cases, non-convergence occurred. 3P results in slightly less accurate coregistration with almost no cases of non-convergence. Both automatic methods (ANTs, Elastix) showed significantly worse scores. ANTs advanced normalization tools, 3P three point alignment, TIE thick inner ear mask coregistration.

Fig. 4

Slicewise comparisons of the 4 examined registration methods and 3D-renderings of the resulting median volumes from 153 human inner ears (from top to bottom: automatic Elastix registration, automatic ANTs registration, semimanual three-point registration (3P), automatic mask-aided TIE-approach). Finer inner ear details such as the semicircular canals are only discernible in the 3P and TIE methods.

Fig. 5

Colocalization scatterplots of each registration method with the target atlas (from left to right: Elastix, ANTs, 3P, TIE). In order to exclude background noise, an inner ear mask was used as an ROI (diagonal line: linear regression fit of colocalizations; horizontal/vertical lines: Costes auto thresholds for channel separation). While exact colocalization (i.e., all data points on the diagonal line) is not possible since the moving image and target image are inherently different, TIE results in the most accurate coregistration, i.e., data points located in close proximity to the diagonal line.