All animal manipulations in this work were performed in accordance with Belgian legislation and the directives set by the Ethical committee on Animal Experimentation of our institution (University of Antwerp, Belgium). Three adult Mongolian gerbils (Meriones unguiculatus), aged between 3 and 6 months, were used. They were housed in cages with food and water ad libitum in our animal facility.
The animals were euthanized using carbon dioxide, followed by a cardiac perfusion with physiological fluid to rinse out all the blood from the gerbil head blood vessels. This step is necessary to allow for OPFOS tomography (as we will explain below). The gerbils are then decapitated and the right temporal bones were isolated. The specimens were reduced in size until only the bulla was left containing the middle and inner ear, cf. Figure . During the harvesting of these bullas, continuous moistening with mist from an ultrasonic humidifier (Bionaire BT-204) was applied to avoid dehydration.
Cross-sectional imaging of bone with X-ray tomography
The first stage of 3D tomographic recording of the ME was achieved using micro-scale X-ray computed tomography. The dissected bullae were enclosed in separate Eppendorf vials, together with a calibration object and a few droplets of physiological fluid at the bottom. In this way, a 100% saturated humid environment was created to avoid dehydration artifacts. Another droplet of fluid was placed in the ear canal—which could help—to distinguish the outline border of the TM shape with the air-filled ME cavity. Water and air have a slightly different X-ray absorption coefficient, so a layer of water on the extremely thin TM can help to reveal its medial shape outline. In previous work, we measured the shape of the eardrum before and after putting fluid on the membrane: Even with a 10 mm water column in the ear canal, no measurable deformation was found with moiré profilometry of 15 μm resolution (Buytaert et al. 2009
). As the droplet of water is less than 3 mm high (inducing a pressure load of 30 Pa), the TM deformation is well below the μCT measurement resolution. The Eppendorf vials (made from polypropylene) are almost X-ray transparent. Especially bone absorbs X-rays well, thus creating a high contrast in transmission recordings. The small calibration objects were custom-made from polyvinyl chloride in our mechanical workshop and possess about the same X-ray absorption properties as thin bone (Gea et al. 2005
). They served as an independent calibration to verify the μCT device specifications.
The vials containing gerbil specimens were scanned at the UGCT scanning facility at Ghent University (www.ugct.ugent.be
) using a custom-built μCT scanner of medium energy (up to 160 keV). The scanner has a directional X-ray tube with a feature recognition capability up to 2 μm (Masschaele et al. 2007
). The scans were performed at a tube voltage of 120 kV (photon energy levels ranging from 0 to 120 keV) and a current of 58 μA. A custom-made vial holder was mounted on a computer-controlled rotation table (MICOS, UPR160F-AIR). For each specimen, a series of 1,000 shadow projections of 1,496
1,880 pixels was recorded covering an object rotation of 360° (or one recording every 0.36°). Reconstruction of the tomographic data volume to serial sections was achieved using the back-projection algorithms of the Octopus software package (Dierick et al. 2004
), resulting in 1,780 reconstructed cross sections of 1,496
1,496 pixels. From these calculated cross sections with an isometric pixel size of 8.5 μm, accurate 3D models of the three ossicles and other bony structures were generated. All three datasets cover a volume of 15.1
12.7 mm (1,780
Cross-sectional imaging of soft tissue with optical tomography
Due to the low X-ray absorption of soft tissue, another tomographic technique was needed: OPFOS microscopy (Voie et al. 1993
). OPFOS was initially developed to image the inner ear cochlea, but it has also been used in ME studies (Voie 2002
; Buytaert and Dirckx 2007
). In the OPFOS method, parallel optical sections through a macroscopic biomedical specimen are created by means of a thin sheet of laser light, and the fluorescence originating from within the cross section of the light sheet with the tissue is recorded in the direction perpendicular to the plane of the laser light. The light emitted by the specimen originates from auto-fluorescence or from staining the specimen with a fluorescent dye. OPFOS images both bone and soft tissue at the same time and in real time, as no (back-projection) calculations are required. It allows region-of-interest (ROI) imaging and has both a high sectioning and a high in-plane resolution. Hence, perfectly and automatically aligned images of virtual cross sections can be obtained. OPFOS scanning was performed at the Laboratory of BioMedical Physics at the University of Antwerp (www.ua.ac.be/bimef
) with a custom-built setup using bi-directional light-sheet illumination (Buytaert and Dirckx 2009
; Buytaert 2010
For OPFOS imaging, an elaborate specimen preparation is needed (Voie 2002
; Buytaert and Dirckx 2009
), as the technique requires the specimens to be perfectly transparent. Before μCT scanning, all blood was removed from the blood vessels, as coagulated blood cannot be made transparent afterward. After μCT recording, a 10% neutral buffered formalin bath was applied. Next, all calcium was removed using 10% EDTA in water solution combined with microwaves. Because of this decalcification, the OPFOS method has to be performed second after μCT X-ray scanning. Then, the specimens were dehydrated using a slowly graded ethanol series, up till 100%. Next, all tissue was refractive index matched using a slowly graded Spalteholz fluid series, again up till 100%. As a result, the specimens become entirely transparent when submerged in pure Spalteholz fluid. Finally, to obtain stronger fluorescence, the specimens are stained with rhodamine B.
Both soft tissue and bone were made transparent and fluorescent; hence, both tissue types are visualized with the technique. We focused on ROI OPFOS imaging of ME ligaments, tendons, and muscles, while images of the (often larger) bony structures are more easily obtained from μCT. Comparison of high-resolution μCT and OPFOS data allows us to distinguish bone from soft tissue in the OPFOS data. Merging of the two datasets generates the complete ME model with all of its functional components accounted for.
The shape of the TM was obtained from the μCT data. The OPFOS technique is able to visualize this extremely thin tissue when performing ROI imaging on a small part of the membrane, cf. Figure . However, to image the membrane full-field with OPFOS, one needs to zoom out and the resolution needed to adequately visualize this thin membrane is lost. Furthermore, the eardrum is prone to preparation artifacts: Because the gerbil specimens went through an extensive procedure of tissue fixation, decalcification, dehydration, and Spalteholz treatment, the extremely thin TM can get deformed. Therefore, the data on eardrum shape are obtained from the CT images, recorded before any specimen processing was applied. X-rays are normally not suited to image soft tissue, especially if it is very thin, like the eardrum. We tried to counter this problem by applying a droplet of physiological fluid through the ear canal on top of the membrane. The medial border of the droplet and eardrum then becomes more easily distinguishable from air in the ME cavity. In this way, the membrane outline will be obtained without deformation and with adequate resolution.
FIG. 2. 2D virtual cross sections delivered by the OPFOS technique. A Tensor tympani tendon reaching down toward malleus. B Incudomallear and incudostapedial articulation. Pixel size 1.5×1.5 μm.
Apart from the specimen preparation, the OPFOS method has another disadvantage as it suffers from stripe artifacts. Opaque regions or areas of less transparency locally reduce the intensity of the laser light sectioning sheet, causing shadow lines or stripes in the rest of the image. This is partially countered by simultaneous dual light-sheet illumination in our setup (Buytaert and Dirckx 2009
). Measuring and analyzing the OPFOS data is very time-consuming; therefore, only one gerbil ear has been processed. On the other hand, the μCT data of all three gerbils were analyzed.
We performed visual observations of the orientation, location, shape, and suspension of the ossicular chain inside opened ME bullae with an operating microscope (Zeiss, OPMI Sensera S7). When 3D computer data, models and results were obtained from μCT or OPFOS with striking features, they were compared to qualitative observations of the real geometry in opened bullae with the operating microscope to verify their interpretation. These experiences gave us the necessary expertise to confirm the 3D model results and conclusions of the present paper. For instance, after a targeted dissection, we could visually confirm that the posterior incudal ligament in gerbil indeed exists as one whole band instead of two separate structures, as we found in our OPFOS data and model.
3D segmentation and reconstruction
After obtaining several series of object cross sections—one μCT set originating from back-projection calculations and several ROI datasets from direct OPFOS recordings—we identified and segmented the relevant structures in all images. The goal of segmentation is to locate objects boundaries, which in turn allows software to build 3D surface meshes by triangulation.
In our case, segmentation was done manually for thousands of sections using the commercial image segmentation and 3D surface mesh generating software package Amira 5.3 (Visage Imaging). Manual segmentation might seem primitive and time-consuming, but using our morphological expertise, manual segmentation delivers better results than purely automated segmentation based on thresholding of gray scale values. The Amira software package uses the marching cubes algorithm for triangulation. It takes eight neighboring voxel locations at a time (forming an imaginary cube), after which the polygon(s) needed to represent the part of the isosurface that passes through this cube are determined. The individual polygons are finally fused into the intended surface. This leads to subvoxel triangulation that easily manages sharp angles. When smoothing or simplification (reduction of the number of triangles) is used, the program takes the “steepness” of the surface into account: Flat surface parts are more reduced than curved parts.
As final result, we end up with triangulated surface meshes for the μCT and OPFOS datasets. These can be further developed into finite-element volume meshes using Amira or other packages. On the website of the Laboratory of BioMedical Physics group, we suggest some powerful and open-source volume generating software, e.g., PreView.
Merging of CT and OPFOS models
All cross sections in a μCT dataset and therefore all models of ME components originating from it are inherently perfectly aligned within the data stack. The OPFOS datasets were focused on the soft tissue by separate ROI recordings. However, parts of the bone are included in the OPFOS ROI recordings as well. The cross sections within each ROI OPFOS data stack are also perfectly aligned, but the resulting mesh models per stack are unrelated to the other OPFOS datasets (because of different ROI zooming and/or other slicing orientation) and unrelated to the CT dataset and models.
To merge the OPFOS data with the μCT data, the μCT dataset was used as a reference. We did not merge the 2D image cross sections, but the 3D mesh models: All partial bone models from ROI OPFOS were three-dimensionally aligned to corresponding parts of the μCT models using an iterative spatial transformation least-squares minimization process of the Amira software package. This process uses the iterative closest point (ICP) algorithm to minimize the difference between two point clouds (e.g., all surface nodes of, respectively, an OPFOS and a μCT mesh model). ICP iteratively revises the spatial transformation (6 degrees of freedom for translation and rotation) needed to minimize the Euclidean distance between the points of two datasets. This concept is referred to as the Procrustes superimposition method: The root mean square (RMS) of the distances between corresponding points of the two surfaces is evaluated. Corresponding point pairs are created by finding the closest point of the reference (μCT) surface mesh for each point of the other (OPFOS) surface mesh. When the two surfaces are identical and perfectly superimposed, the RMS of all corresponding point distances will be zero. In the case of the OPFOS versus the μCT stapes model for instance, we obtained a root mean square difference of 17 μm (or two μCT voxels). After obtaining such a good match between the OPFOS and μCT bone model, we applied the same spatial transformation to the OPFOS soft tissue mesh(es) from that OPFOS dataset. In this way, all OPFOS datasets were combined with μCT data into one model.