Combining Interactive and Automatic Volume Registration Techniques in Tomviz
Patrick Avery, Alessandro Genova, Matthew M. McCormick, Yu‐chen Karen Chen‐Wiegart · Microscopy and Microanalysis · 2024
Tomviz is a cross-platform, open source tool for the visualization and analysis of tomography data [1]. Built on top of the industry-leading open source visualization platforms ParaView [2] and VTK [3], Tomviz provides state-of-the-art rendering techniques and a broad collection of tools for image processing such as tilt image alignment, tomographic reconstruction, histograms, filters, user-customized Python scripts, and many researcher-contributed algorithms. Image registration is the process of transforming the pixels of an image to align its contained data to similar data in another image. It often involves a "moving image,” which is transformed and aligned to a stationary “fixed image.” A volume, which is essentially a 3D image, can also be registered, with the additional dimension producing a more challenging problem. When comparing similar datasets, misaligned volumes are common since they are often measured at different times and conditions (Fig. 1 provides an example of two misaligned X-ray nano-tomography datasets). But to perform accurate analytical comparisons between datasets, aligning the volumes through registration is often necessary. Interactive/manual volume registration is the process of manually translating, rotating, and scaling one volume to align it with another. Automatic registration, which can require less human intervention, is typically preferred, but noisy or highly misaligned volumes can cause automatic registration to not succeed, thus necessitating manual registration. Without intuitive interaction to aid in the selection of transformation parameters, however, manual registration can easily become a highly time-consuming trial-and-error process. In Tomviz, a new “Manual Manipulation” operator facilitates manual registration through familiar visualization scene interactions (Fig. 2). Mouse interactions with the selected volume are nearly identical to familiar camera manipulations (rotating, translating, and zooming) that are performed when interacting with the scene, except that all interactions are recorded for use in the registration afterward. Internally, the interactions simply transform the volume’s “visualization element” to provide a fast, interactive, and accurate preview of the transformed volume, without applying the more time-consuming voxel transformations in the actual registration that follows. After the “moving” volume appears aligned to the “fixed” volume and the transformations are accepted, the moving volume’s voxels are resampled through rotating, shifting, and scaling to produce a new volume aligned with the reference volume. Optionally, the moving volume’s voxels are also cropped, padded, and subsampled so that the shape/dimensions of the moving volume exactly match that of the reference volume as well. Such interactive determination of transformation parameters significantly simplifies and improves the manual volume registration process. When the two volumes are satisfactorily “roughly aligned”, a new automatic “Registration” operator in Tomviz may be used to complete the alignment. The automatic volume registration in Tomviz utilizes the open source ITKElastix project [4, 5]. ITKElastix provides an Insight Toolkit (ITK) [6, 7] interface to elastix [8], a toolbox of algorithms for volume registration techniques. Some example options are shown in Fig. 3. The use of ITKElastix provides access to a broad collection of registration algorithms, which may be utilized through the highly customizable Python operator scripts in Tomviz (the current automatic registration operator script being a great starting point). Automatic registration can sometimes require parameter tuning to obtain the best results, but it can often be used to align the volumes very closely. Fig. 4 shows that after automatic registration, the two volumes are very well aligned. Such alignment is then the foundation for the comparative analysis to be done afterward. Two unaligned volumetric datasets of a Ni–20Cr microwire measured through X-ray nano-tomography [9]. The first view is the pristine form of the microwire, before any corrosion has occurred. The second view is a corroded form of the microwire after reacting with KCl–MgCl2 molten salt at 800 °C for 5 minutes. The third view is a combined view with both datasets present, where the misalignment may be easily seen. Manual volume registration of the two datasets. Mouse interactions with the selected volume are nearly identical to the already familiar camera manipulations (rotating, translating, and zooming) performed when interacting with the scene, except that all interactions are recorded for use in the registration afterward. The red outline indicates the changes to the volume’s “visualization element” boundaries before the registration is applied. “Align voxels with reference?” ensures the voxels are aligned with the reference dataset’s voxels by performing cropping, padding, and subsampling. After “OK” is clicked, the voxels are transformed according to the recorded user interactions. Example ITKElastix registration parameters utilized for automatic registration. Final rendering of both volumetric datasets aligned with one other, after both manual and automatic registration.