CAT: a computational anatomy toolbox for the analysis of structural MRI data

1The study of the human brain using neuroimaging methods is still in its infancy, but rapid technical advances in image acquisition and processing are enabling ever more refined characterizations of its micro- and macro-structure. Enormous efforts, for example, have been made to map differences between groups (e.g., young vs. old, diseased vs. healthy, male vs. female), to capture changes over time (e.g., from infancy to old age, in the framework of neuroplasticity, as a result of a clinical intervention), or to assess correlations of brain attributes (e.g., measures of length, volume, shape) with behavioral, cognitive, or clinical parameters. Popular neuroimaging software packages include tools for analysis and visualization, such as SPM (RRID:SCR_007037), FreeSurfer (RRID:SCR_001847), the Human Connectome Workbench, FSL (RRID:SCR_002823), BrainVISA, CIVET, or the LONI tools, just to name a few.

2SPM (short for Statistical Parametric Mapping) is one of the most frequently used software packages, which works with MATLAB (RRID:SCR_001622) as well as Octave. Its library of accessible and editable scripts provides an ideal basis to extend the repertoire of preprocessing and analysis options. Over the years, SPM has inspired developers to create powerful tools that use SPM’s functionality and interface. These tools are more than just extensions of SPM, offering a comprehensive range of cutting-edge options across the whole analysis spectrum, from the initial data processing to the final visualization of the statistical effects.

3One such tool is CAT (short for Computational Anatomy Toolbox). CAT constitutes a significant step forward in the field of human brain mapping by adding sophisticated methods to process and analyze structural brain magnetic resonance imaging (MRI) data using voxel-, surface-, and region-based approaches. CAT is available as a collection of accessible scripts, with an intuitive user interface, and uses the same batch editor as SPM, which allows for a seamless integration with SPM workflows and other toolboxes, such as Brainstorm and ExploreASL. Not only does this enable beginners and experts to run complex state-of-the-art structural image analyses within the SPM environment, but it will also provide advanced users as well as developers the much-appreciated option to incorporate a wide range of functions in their own customized workflows and pipelines.

4CAT12 is the current version of the CAT software and runs in MATLAB (MathWorks) or as a standalone version with no need for a MATLAB license. It was originally designed to work with SPM12 and is compatible with MATLAB versions 7.4 (R2007a) and later. No additional software or toolbox is required. The latest version of CAT can be downloaded here:. The precompiled standalone version for Windows, Mac, or Linux operating systems can be downloaded here:. All steps necessary to install and run CAT are documented in the user manual and in the complementary online help, which can be accessed directly via CAT’s help functions. The CAT software is free but copyrighted and distributed under the terms of the GNU General Public License, as published by the Free Software Foundation.

5CAT can be started through SPM, from the MATLAB command window, from a shell, or as a standalone version. Except when called from the command shell (CAT is fully scriptable), a user interface will appear (see Fig. 1), allowing easy access to all analysis options and most additional functions. In addition, a graphical output window will display the interactive help to get started. This interactive help will be replaced by the results of the analyses (i.e., in that same window) but can always be called again via the user interface.

6CAT’s processing pipeline (see Fig. 2) contains 2 main streams: (i) voxel-based processing for voxel-based morphometry (VBM) and (ii) surface-based processing for surface-based morphometry (SBM). The former is a prerequisite for the latter, but not the other way round. Both processing streams can be extended to include additional steps for (iii) region-based processing and region-based morphometry (RBM).

7Voxel-based processing steps can be roughly divided into a module for tissue segmentation, followed by a module for spatial registration.

8Tissue Segmentation: The process is initiated by applying a spatially adaptive nonlocal means (SANLM) denoising filter, followed by SPM’s standard unified segmentation. The resulting output serves as a starting point for further optimizations and CAT’s tissue segmentation steps: first, the brain is parcellated into the left and right hemispheres, subcortical areas, ventricles, and cerebellum. In addition, local white matter hyperintensities are detected (to be later accounted for during the spatial registration and the optional surface processing). Second, a local intensity transformation is performed to reduce the effects of higher gray matter intensities in the motor cortex, basal ganglia, and occipital lobe, which are influenced by varying degrees of myelination. Third, an adaptive maximum a posteriori (AMAP) segmentation is applied, which does not require any a priori information on the tissue probabilities. The AMAP segmentation also includes a partial volume estimation. Figure 3A provides information on the accuracy of CAT’s tissue segmentation.

9Spatial Registration: Geodesic Shooting is used to register the individual tissue segments to standardized templates in the ICBM 2009c Nonlinear Asymmetric space (MNI152NLin2009cAsym), hereafter referred to as MNI space. While MNI space is also used in many other software packages, enabling cross-study comparisons, users may also choose to use their own templates. Figure 3B provides information on the accuracy of CAT’s spatial registration.

10VBM is applied to investigate the volume (or local amount) of a specific tissue compartment—usually gray matter. VBM incorporates different processing steps: (i) tissue segmentation and (ii) spatial registration, as detailed above, and in addition, (iii) adjustments for volume changes due to the registration (modulation) as well as (iv) convolution with a 3-dimensional (3D) Gaussian kernel (spatial smoothing). As a side note, the modulation step results in voxel-wise gray matter volumes that are the same as in native space (i.e., before spatial registration) and not corrected for brain size yet. To remove effects of brain size, users have at least 2 options: (i) calculating the total intracranial volume (TIV) and including TIV as a covariate in the statistical model or (ii) selecting “global scaling” (see second-level options in SPM). The latter is recommended if TIV is linked with (i.e., not orthogonal to) the effect of interest (e.g., sex), which can be tested (see “Design orthogonality” in SPM).

11The optional surface-based processing comprises a series of steps that can be roughly divided into a module for surface creation, followed by a module for surface registration.

12Surface Creation: Fig. 3 illustrates the surface creation step in CAT for data obtained on scanners with different field strengths (1.5, 3.0, and 7.0 Tesla). CAT uses a projection-based thickness method, which estimates the initial cortical thickness and initial central surface in a combined step, while handling partial volume information, sulcal blurring, and sulcal asymmetries, without explicit sulcus reconstruction. After this initial step, topological defects (i.e., anatomically incorrect connections between gyri or sulci) are repaired using spherical harmonics. The topological correction is followed by a surface refinement, which results in the final central, pial, and white surface meshes. In the last step, the final pial and white matter surfaces are used to refine the initial cortical thickness estimate using the FreeSurfer thickness metric. Alternatively, the final central surface can be used to calculate metrics of cortical folding, as described under “Surface-based morphometry (SBM).”

13Surface Registration: The resulting individual central surfaces are registered to the corresponding hemisphere of the FreeSurfer FsAverage template. During this process, the individual central surfaces are spherically inflated with minimal distortions, and a one-to-one mapping between the folding patterns of the individual and template spheres is created by a 2-dimensional (2D) version of the DARTEL approach. Figure 3D provides information on the accuracy of CAT’s surface registration.

14SBM can be used to investigate cortical thickness or various parameters of cortical folding. The measurement of “cortical thickness” captures the width of the gray matter ribbon as the distance between its inner and outer boundary at thousands of points (see Fig. 4). To obtain measurements of “cortical folding,” the user has a variety of options in CAT, ranging from Gyrification to Sulcal Depth to Cortical Complexity to the Surface Ratio, as explained and illustrated in Fig. 4. Similar to VBM, SBM incorporates a series of different steps: (i) surface creation and (ii) surface registration, as detailed above, and (iii) spatial smoothing. As a side note, since the measurements in native space are mapped directly to the template during the spatial registration, no additional modulation (as in VBM) is needed to preserve the individual differences. In contrast to VBM, SBM does not require brain size corrections because cortical thickness and cortical folding are not closely associated with total brain volume (unlike gray matter volume).

15In addition to voxel- or point-wise analyses via VBM or SBM, CAT provides an option to conduct regional analyses via region-based morphometry (RBM). For this purpose, the processing steps under voxel-based processing (surface-based processing, respectively) should be applied and followed by automatically calculating regional measurements. This is achieved by working with regions of interest (ROIs), defined using standardized atlases. The required atlases are provided in CAT (see Supplementary Table S1 and Supplementary Table S2), but users can also work with their own atlases.

16Voxel-based ROIs: The volumetric atlases available in CAT have been defined on brain templates in MNI space and may be mapped to the individual brains by using the spatial registration parameters determined during voxel-based processing. Volumetric measures, such as regional gray matter volume, can then be calculated for each ROI in native space.

17Surface-based ROIs: The surface atlases available in CAT are supplied on the FsAverage surface and can be mapped to the individual surfaces by using the spherical registration parameters determined during the surface-based processing. Surface-based measures, such as cortical thickness or cortical folding, are then calculated for each ROI in native space.

18CAT allows processing streams to be distributed to multiple processing cores, to reduce processing time. For example, CAT’s analysis of 50 subjects (see “Example application”), leveraging the inbuilt parallel processing capabilities on 4 cores, required 7 hours of processing time when analyzing 1 image per subject (cross-sectional stream) and 18 hours when processing 3 images per subject (longitudinal stream) for the entire sample. Application of all available workflows for a single T1-weighted image takes around 35 minutes, as timed on an iMac with Intel Core i7 with 4 GHz and 32 GB RAM using MATLAB version 2017b, SPM12 version r7771, and CAT12.8 version r1945.

19CAT’s performance has been thoroughly tested by evaluating its accuracy, sensitivity, and robustness in comparison to other tools frequently used in the neuroimaging community. For this purpose, we applied CAT and analyzed real data (see “Example application”) as well as simulated data generated from BrainWeb. The evaluation procedures are detailed in Supplementary Note 1 and Supplementary Note 2; the outcomes are presented in Supplementary Fig. S1 and Supplementary Fig. S2. CAT proved to be accurate, sensitive, reliable, and robust, outperforming other common neuroimaging tools.

20Aside from offering a standard pipeline for cross-sectional analyses, CAT has specific longitudinal pipelines that ensure a local comparability both across subjects and across time points within subjects. Compared to the cross-sectional pipeline, these longitudinal pipelines render analysis outcomes more accurate when mapping structural changes over time. The user can choose between 3 different longitudinal pipelines: the first one for analyzing brain plasticity (over days, weeks, months), the second one for analyzing brain development (over months and years), and the third one for brain aging (over months, years, decades). For more details, refer to Supplementary Note 3.

21CAT introduces a retrospective quality control framework for the empirical quantification of essential image parameters, such as noise, intensity inhomogeneities, and image resolution (all of these can be impacted, for example, by motion artifacts). Separate parameter-specific ratings are provided as well as a handy overall rating. Moreover, image outliers can be easily identified, either directly based on the aforementioned indicators of the image quality or by calculating a z-score determined by the quality of the image processing as well as by the anatomical characteristics of each brain. For more details, refer to Supplementary Note 4.

22CAT allows the user to map voxel-based values (e.g., quantitative, functional, or diffusion parameters) to individual brain surfaces (i.e., pial, central, and/or white matter) for surface-based analyses. The integrated equi-volume model also considers the shift of cytoarchitectonic layers caused by the local folding. Optionally, CAT also allows mapping of voxel values at multiple positions along the surface normal at each node—supporting a layer-specific analysis of ultra-high resolution functional MRI data. For more details, refer to Supplementary Note 5.

23CAT comes with its own threshold-free cluster enhancement (TFCE) toolbox and provides the option to apply TFCE in any statistical second-level analysis in SPM, for both voxel-based and surface-based analyses. It can also be employed to analyze functional MRI (fMRI) or diffusion tensor imaging (DTI) data. A particularly helpful feature of the TFCE toolbox is that it automatically recognizes exchangeability blocks and potential nuisance parameters from an existing statistical design in SPM. For more details, refer to Supplementary Note 4.

24CAT allows a user to generate graphs and images, which creates a solid basis to explore findings as well as to generate ready-to-publish figures according to prevailing standards. More specifically, it includes 2 distinct sets of tools to visualize results: the first set prepares both voxel- and surface-based data for visualization by providing options for thresholding the default SPM T-maps or F-maps and for converting statistical parameters (e.g., T-maps and F-maps into p-maps). The second set of tools visualizes the data offering the user ample options to select from different brain templates, views, slices, significance parameters, significance thresholds, color schemes, and so on (see Fig. 5).

25To demonstrate an application of CAT, we investigated an actual dataset focusing on the effects of Alzheimer’s disease on brain structure. More specifically, we set out to compare 25 patients with Alzheimer’s disease and 25 matched controls. We applied (i) a VBM analysis focusing on voxel-wise gray matter volume, (ii) an RBM analysis focusing on regional gray matter volume (i.e., a voxel-based ROI analysis), (iii) a surface-based analysis focusing on point-wise cortical thickness, and (iv) an RBM analysis focusing on regional cortical thickness (i.e., a surface-based ROI analysis). Given the wealth of literature on Alzheimer’s disease, we expected atrophy in gray matter volume and cortical thickness in patients compared to controls, particularly in regions around the medial temporal lobe and the default mode network. In addition to distinguishing between the 4 morphological measures (i–iv), all analyses were conducted using both cross-sectional and longitudinal streams in CAT. Overall, we expected that longitudinal changes would manifest in similar brain regions to cross-sectional group differences but that cross-sectional effects would be more pronounced than longitudinal effects. The outcomes of this example analysis are presented and discussed in the next section.

26As shown in Fig. 6, all 4 cross-sectional streams—investigating voxel-based gray matter volume, regional gray matter volume, point-wise thickness, and regional thickness—revealed widespread group differences between patients with Alzheimer’s disease (AD) and matched controls. Overall, the effects were comparable between cross-sectional and longitudinal streams, but the significant clusters were more pronounced cross-sectionally (note the different thresholds cross-sectionally and longitudinally).

27More specifically, using VBM, significantly smaller voxel-wise gray matter volumes were observed in patients with AD compared to controls, particularly in the medial and lateral temporal lobes and within regions of the default mode network (Fig. 6A, top). Similarly, the longitudinal follow-up revealed a significantly stronger gray matter volume loss in patients compared to controls, with effects located in the medial temporal lobe as well as the default mode network (Fig. 6A, bottom). The voxel-based ROI analysis resulted in a significance pattern similar to the VBM study, with particularly pronounced group differences in the temporal lobe that extended into additional brain areas, including those comprising the default mode network (Fig. 6B, top). Again, the longitudinal analysis yielded similar but less pronounced findings than the cross-sectional analysis, although longitudinal effects were stronger than in the VBM analysis (Fig. 6B, bottom).

28Using SBM, the point-wise cortical thickness analysis yielded a pattern similar to the VBM analysis with significantly thinner cortices in patients, particularly in the medial and lateral temporal lobe and within regions of the default mode network (Fig. 6C, top). Just as in the VBM analysis, significant clusters were widespread and reached far into adjacent regions. Again, the results from the longitudinal stream were less widespread and significant than the results from the cross-sectional stream (Fig. 6C, bottom). Finally, the surface-based ROI analysis largely replicated the local findings from the SBM analysis (Fig. 6D, top/bottom).

29Overall, the results of all analysis streams corroborate prior findings in the Alzheimer’s disease literature, particularly the strong disease effects within the medial temporal lobe and regions of the default mode network. Furthermore, the comparable pattern across measures suggests a considerable consistency between available morphometric options, even if gray matter volume and cortical thickness are biologically different and not perfectly related.

30As shown in Supplementary Fig. S1 and Supplementary Fig. S2, CAT12 proved to be accurate, sensitive, reliable, and robust, outperforming other common neuroimaging tools. Similar conclusions have been drawn in independent evaluations testing 1 or more software in comparison with CAT12. For example, Guo et al. evaluated the repeatability and reproducibility of brain volume measurements using FreeSurfer, FSL-SIENAX, and SPM and highlighted the reliability of CAT12. Similarly, CAT12 emerged as a robust option when demonstrating that the choice of the processing pipeline influences the location of neuroanatomical brain markers. Last but not least, Khlif et al. compared the outcomes of CAT12’s automated segmentation of the hippocampus with those achieved based on manual tracing and demonstrated that both approaches produced comparable hippocampal volume.

31In addition, numerous evaluations suggest that CAT12 performs at least as well as other common neuroimaging tools and, as such, offers a valuable alternative. For example, Tavares et al. conducted a VBM study and concluded that the segmentation pipelines implemented in CAT12 and SPM12 provided results that are highly correlated and that the choice of the pipeline had no impact on the accuracy of any brain volume measure. Along the same lines, but for SBM, Ay et al. reported that CAT12 and FreeSurfer produced equally valid results for parcel-based cortical thickness calculations. de Fátima Machado Dias et al. addressed the issue of reproducibility and observed that cortical thickness measures using CAT12 and FreeSurfer were comparable at the individual level. Moreover, Seiger et al. conducted a study in patients with Alzheimer’s disease and healthy controls, in which CAT12 and FreeSurfer provided consistent cortical thickness estimates and excellent test–retest variability scores. Velázquez et al. supported these findings when comparing CAT12 and FreeSurfer with 3 voxel-based methods in a test–retest analysis and clinical application. Finally, Righart et al. compared volume and surface-based cortical thickness measurements in multiple sclerosis and emphasized CAT12’s consistent performance.

32These collective findings from multiple studies support the notion that CAT is a robust and reliable tool for both VBM and SBM analyses, producing results that are comparable to and, in some cases, superior to other established neuroimaging software.

33CAT is suitable for desktop and laptop computers as well as high-performance clusters. It is fully integrated into the SPM environment within MATLAB but also allows process execution directly from the command shell, without having to start SPM. CAT can also run without a MATLAB license by using the stand-alone version or by using Octave instead of MATLAB. In terms of performance, CAT allows for ultra-fast processing and analysis and also is more sensitive in detecting significant effects compared to other common tools used by the neuroimaging community. Moreover, it better handles varying levels of noise and signal inhomogeneities. Furthermore, CAT is easy to integrate with non-SPM software packages and also supports the Brain Imaging Data Structure (BIDS) standards. Therefore, CAT is ideally suited to process not only small datasets (as demonstrated in the example application) but also big datasets, such as samples of the UK Biobank or ENIGMA. Finally, while CAT is currently targeted at structural imaging data, some features (e.g., high-dimensional spatial registration or mapping onto the cortical surface) may also be used for the analysis of functional, diffusion, or quantitative MRI or EEG/MEG data.

34Data for the application example were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. The ADNI (RRID:SCR_003007) was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early AD. For up-to-date information, see.

35For the purpose of the current study, we compiled a sample of 50 subjects with 3D T1-weighted structural brain images from the ADNI database. Specifically, we randomly selected the first 25 subjects (16 males/9 females) classified as patients with AD (mean age 75.74 ± 8.14 years; mean Minimal Mental Status Examination [MMSE] score: 23.44 ± 2.04) and matched them for sex and age with 25 healthy controls (mean age 76.29 ± 3.90 years; mean MMSE: 28.96 ± 1.24). Informed consent was obtained from all research participants. All subjects had brain scans at baseline (first scan at enrollment) and at 2 follow-up visits, at 1 year and 2 years after the first scan. All brain images were acquired on 1.5 Tesla scanners (Siemens, General Electric, Philips) using a 3D T1-weighted sequence with an in-plane resolution between 0.94 and 1.25 mm and a slice thickness of 1.2 mm.

36All T1-weighted data were processed using CAT12 following the cross-sectional (or longitudinal, respectively) processing stream for VBM, SBM (cortical thickness), and ROI analyses (see Fig. 2) according to the descriptions provided under “Computational morphometry.” For each subject, only their first time point was included in the cross-sectional stream, whereas all 3 time points were included in the longitudinal stream. The processing streams for the VBM analysis resulted in modulated and registered gray matter segments, which were smoothed using a 6-mm Gaussian kernel. The image-processing streams for the SBM analysis resulted in the registered point-wise cortical thickness measures, which were smoothed using a 12-mm Gaussian kernel. The voxel-based ROI analysis used the Neuromorphometrics atlas (RRID:SCR_005656) to calculate the regional gray matter volumes; the surface-based ROI analysis employed the DK40 atlas to calculate regional cortical thickness.

37For each variable of interest—voxel-wise gray matter volume, regional gray matter volume, point-wise cortical thickness, and regional cortical thickness—the dependent measures (e.g., the registered, modulated, and smoothed gray matter segments for voxel-wise gray matter) were entered into the statistical model. For the cross-sectional stream, group (patients with AD vs. controls) was defined as the independent variable. For the longitudinal stream, the interaction between group and time was defined as the independent variable, whereas subject was defined as a variable of no interest. For the VBM and the voxel-based ROI analyses, data were corrected for TIV using “global scaling” (because TIV correlated with group, the effect of interest). Since cortical thickness does not scale with brain size, no corrections for TIV were applied for the SBM and the surface-based ROI analyses. For the cross-sectional analysis, we additionally included age as a nuisance parameter.

38For the VBM and SBM analyses, results were corrected for multiple comparisons by applying TFCE and controlling the family-wise error at P ≤ 0.001 (cross-sectional) and P ≤ 0.05 (longitudinal). For the voxel-based and surface-based ROI analyses, results were corrected for multiple comparisons by controlling the false discovery rate at q ≤ 0.001 (cross-sectional) and q ≤ 0.05 (longitudinal). All statistical tests were 1-tailed given our a priori hypothesis that patients with AD have less gray matter at baseline and a larger loss of gray matter over time.

39The outcomes of the VBM and voxel-based ROI analyses were overlaid onto orthogonal sections of the average brain that was created from the spatially registered T1-weighted images of the study sample (n = 50); the outcomes of the SBM and surface-based ROI analyses were projected onto the FsAverage surface.

40Project name: Computational Anatomy Toolbox

41Project homepage:

42Software documentation:

43Operating system(s): Platform independent (MacOS, Linux, Windows)

44Programming language: MATLAB, C

45Other requirements: MATLAB (7.4 or newer)

46License: GPL 2.0

47RRID:SCR_019184