CTseg: CT Brain Segmentation Tool

• CTseg is a brain CT segmentation and spatial normalisation tool that adapts SPM’s unified segmentation to CT imaging using a probabilistic generative model.
• It produces SPM-compatible outputs including tissue maps, deformation fields, and brain volume estimates, enabling integration with established neuroimaging workflows.
• Validation shows CTseg outperforms MRI-based SPM segmentation in key metrics such as Dice scores and TBV agreement while operating without preprocessing.

CTseg is a freely available software tool for brain CT segmentation, spatial normalisation, and volumetrics. It builds on the Multi-Brain generative modelling framework and provides a CT-specific pipeline that produces tissue maps, deformation fields, and brain volume estimates in the same format as SPM’s unified segmentation, thereby extending SPM’s established analysis chain from MRI to CT. CTseg is designed for routine hospital CT scans and, in deployment, does not require preprocessing or resampling. Its validation used paired MR/CT head scans and assessed segmentation accuracy against an MRI-derived silver standard, spatial normalisation consistency, volumetric agreement, and downstream sex-classification performance from normalised tissue maps (Brudfors, 6 May 2026).

1. Conceptual scope and problem setting

CTseg addresses a specific neuroimaging problem: probabilistic tissue segmentation and diffeomorphic spatial normalisation of brain CT, with native-space derivation of total brain volume and total intracranial volume. The latent tissue label at voxel $i$, denoted $S_i$, takes values in $\{1,\dots,6\}$, corresponding to GM, WM, CSF, bone, soft tissue, and background. The observed image is $I=\{I_i\}$, the atlas prior is $A=\{a_{i,j}\}$, the deformation field is $\phi$, the smooth multiplicative bias field is $B(x)$, and the Gaussian-mixture parameters are $\theta=\{\mu_k,\Sigma_k,\pi_k\}$ (Brudfors, 6 May 2026).

The tool is positioned as a CT analogue of SPM’s MRI-oriented unified segmentation. That positioning is operational rather than merely conceptual: the outputs are SPM-compatible, the deformation fields can be propagated through established SPM workflows, and the toolbox is intended for routine hospital CT scans without skull-stripping, bias correction, orientation harmonisation, or resampling at deployment. CTseg had already been adopted in clinical research spanning stroke, dementia, and brain morphometry, but its 2026 validation addressed the absence of a systematic evaluation against an independent reference standard (Brudfors, 6 May 2026).

A recurrent misconception is that CTseg is simply MRI-style unified segmentation run on CT. The baseline in its validation explicitly applied SPM12 unified segmentation with MRI priors directly to CT using $\texttt{biasreg}=10$, precisely to test that assumption. CTseg was introduced as a CT-specific alternative rather than a parameter tweak of the MRI model.

2. Probabilistic generative model

CTseg adopts a generative formulation in which tissue segmentation, bias-field estimation, and atlas-to-subject alignment are jointly coupled. Each tissue class $j$ is represented by one or more Gaussian components $S_i$0, yielding the conditional likelihood

$S_i$1

where $S_i$2 denotes the set of Gaussian components assigned to class $S_i$3, and $S_i$4 is the bias field at voxel $S_i$5 (Brudfors, 6 May 2026).

The atlas prior is warped through the deformation field. For voxel $S_i$6,

$S_i$7

and therefore

$S_i$8

The joint posterior over segmentation, deformation, bias field, and mixture parameters is proportional to

$S_i$9

After taking the negative log, CTseg optimises the EM-style energy

$\{1,\dots,6\}$0

In this formulation, $\{1,\dots,6\}$1 regularises the deformation field, $\{1,\dots,6\}$2 penalises high-frequency bias structure, and $\{1,\dots,6\}$3 imposes Gaussian priors on $\{1,\dots,6\}$4, $\{1,\dots,6\}$5, and $\{1,\dots,6\}$6 (Brudfors, 6 May 2026).

This model places CTseg in the family of probabilistic atlas-based segmentation methods rather than discriminative CNN pipelines. A plausible implication is that its design prioritises compatibility with deformation-based morphometry and volumetric quantification alongside voxelwise tissue labelling.

3. Unified-segmentation workflow and optimisation

CTseg mirrors SPM’s unified segmentation but replaces MRI-oriented assumptions with a CT-specific atlas, bias-field model, and diffeomorphic Multi-Brain registration model. Deployment is deliberately minimal: the software accepts raw CT with any orientation, voxel size, and field-of-view, and requires no resampling, skull-stripping, or prior bias correction (Brudfors, 6 May 2026).

Bias-field estimation is expressed as a low-dimensional spline or DCT expansion:

$\{1,\dots,6\}$7

with penalty

$\{1,\dots,6\}$8

Given a current deformation field, CTseg computes Gaussian-component posteriors

$\{1,\dots,6\}$9

then pools them into class probabilities $I=\{I_i\}$0 by summing over $I=\{I_i\}$1 (Brudfors, 6 May 2026).

Spatial normalisation is diffeomorphic. The deformation is parameterised as

$I=\{I_i\}$2

where $I=\{I_i\}$3 is a smooth velocity field represented in a DCT-like basis. Its regularisation is

$I=\{I_i\}$4

with $I=\{I_i\}$5 a differential operator such as bending energy. The registration objective couples warped atlas priors to posterior tissue maps:

$I=\{I_i\}$6

Optimisation proceeds by iterative EM. Initialisation uses $I=\{I_i\}$7, $I=\{I_i\}$8, and $I=\{I_i\}$9 from the pre-learned atlas. The E-step computes $A=\{a_{i,j}\}$0 and $A=\{a_{i,j}\}$1; the M-step updates $A=\{a_{i,j}\}$2, then optimises $A=\{a_{i,j}\}$3 and $A=\{a_{i,j}\}$4. Reported convergence occurs after approximately 100 EM iterations, corresponding to about 2–3 minutes per scan on CPU (Brudfors, 6 May 2026).

4. Volumetrics, outputs, and software interoperability

After segmentation, CTseg computes native-space volumetrics by summing tissue probabilities. If $A=\{a_{i,j}\}$5 is voxel volume and $A=\{a_{i,j}\}$6 for $A=\{a_{i,j}\}$7, then

$A=\{a_{i,j}\}$8

and

$A=\{a_{i,j}\}$9

An intracranial mask $\phi$0 is first applied, restricting the sum to voxels for which $\phi$1 (Brudfors, 6 May 2026).

The software accepts a single non-contrast head CT in NIfTI or ANALYZE format, with no orientation or voxel-size constraints. Its outputs are explicitly SPM-compatible and include native-space tissue probability maps $\phi$2 through $\phi$3 for GM through background, the bias estimate $\phi$4, forward and inverse deformations $\phi$5 and $\phi$6, modulated warped tissue maps $\phi$7 through $\phi$8, the tissue-modulated image $\phi$9, the skull-stripped image $B(x)$0, and a CSV summary containing TBV and TIV (Brudfors, 6 May 2026).

Optional smoothing and modulation are available for VBM-style workflows. The provided interfaces include an SPM-12-style MATLAB batch interface and a Nipype wrapper. This suggests that CTseg was designed not only as a standalone segmentation utility but also as a drop-in component for established neuroimaging pipelines.

5. Validation framework and empirical performance

The validation dataset comprised 59 paired CT/MRI scans from SynthRAD2025, preprocessed to a common grid for validation only. Segmentation accuracy against the MRI-derived silver standard was assessed using Dice,

$B(x)$1

the 95th-percentile Hausdorff distance (HD95), and ASSD. Spatial normalisation quality was evaluated using group-average sharpness, defined through the mean image $B(x)$2 and its gradient magnitude $B(x)$3, mean voxelwise coefficient of variation $B(x)$4 within brain, and a warp-validation analysis in which MRI tissue maps were warped using CT-based and MRI-based deformations. Volumetric agreement used ICC(3,1), Pearson $B(x)$5, and Bland–Altman analysis. Downstream predictive validation used modulated warped GM/WM/CSF maps, Jacobian scaling, 8 mm FWHM smoothing, and a Gaussian-process classifier with linear kernel in PRoNTo under 10-fold cross-validation, reporting ROC AUC and balanced accuracy (Brudfors, 6 May 2026).

For segmentation, CTseg significantly outperformed the MRI-based SPM baseline applied directly to CT. Median Dice values improved from $B(x)$6 to $B(x)$7 in GM, from $B(x)$8 to $B(x)$9 in WM, and from $\theta=\{\mu_k,\Sigma_k,\pi_k\}$0 to $\theta=\{\mu_k,\Sigma_k,\pi_k\}$1 in CSF; all corresponding $\theta=\{\mu_k,\Sigma_k,\pi_k\}$2-values were $\theta=\{\mu_k,\Sigma_k,\pi_k\}$3. Surface distances were also significantly lower for CTseg in GM, WM, and CSF, with the exception of CSF HD95 (Brudfors, 6 May 2026).

For spatial normalisation, warping MRI tissue maps by each CT deformation showed higher Dice for CTseg in GM and WM, specifically $\theta=\{\mu_k,\Sigma_k,\pi_k\}$4 versus $\theta=\{\mu_k,\Sigma_k,\pi_k\}$5 for GM and $\theta=\{\mu_k,\Sigma_k,\pi_k\}$6 versus $\theta=\{\mu_k,\Sigma_k,\pi_k\}$7 for WM, both with $\theta=\{\mu_k,\Sigma_k,\pi_k\}$8; CSF was identical at approximately $\theta=\{\mu_k,\Sigma_k,\pi_k\}$9. ASSD also favoured CTseg. For volumetrics, TBV agreement with MRI was stronger for CTseg, with ICC $\texttt{biasreg}=10$0 versus $\texttt{biasreg}=10$1 for the baseline. TIV agreement was comparable, with ICC $\texttt{biasreg}=10$2 for CTseg versus $\texttt{biasreg}=10$3 for the baseline. In the predictive normalisation test, CTseg-based CT achieved AUC $\texttt{biasreg}=10$4, compared with AUC $\texttt{biasreg}=10$5 for the SPM–CT baseline and AUC $\texttt{biasreg}=10$6 for SPM–MR as silver reference. Median runtime per subject was $\texttt{biasreg}=10$7 s for SPM–MR, $\texttt{biasreg}=10$8 s for SPM–CT, and $\texttt{biasreg}=10$9 s for CTseg on a single CPU (Brudfors, 6 May 2026).

These results delimit CTseg’s strengths with some precision. It yielded better segmentation and normalisation than direct MRI-prior transfer to CT, stronger TBV agreement, and comparable TIV agreement, but it did not uniformly dominate every endpoint; the TIV ICC remained similar to the baseline and runtime was longer.

6. Position within CT segmentation research

CTseg belongs to a different methodological lineage from deep-learning systems developed for organ-at-risk segmentation in radiotherapy. In head-and-neck OAR segmentation, for example, SegReg uses a two-stage cascade in which ElasticSyN first registers MRI to CT, after which CT and registered MRI are concatenated as a two-channel volume and processed by a 3D nnU-Net. On HaN-Seg, this multimodal design improved the CT-only baseline from $j$0 to $j$1 in mDSC and from $j$2 to $j$3 in mIoU, with especially large gains for small OARs such as the cochlea, optic nerves, and lacrimal glands (Zhang et al., 2023). CTseg does not pursue that objective: it is a brain CT tissue-segmentation and normalisation framework, not a multimodal OAR segmenter.

Likewise, CTseg is distinct from thoracic CT benchmarks such as SegTHOR, a dataset of 60 thoracic CT volumes with manual labels for heart, aorta, trachea, and esophagus. Baseline 2D U-Net variants on SegTHOR reported strong performance for large organs and more difficulty for smaller, lower-contrast structures, with a simplified U-Net reaching Dice values of $j$4 for esophagus, $j$5 for trachea, $j$6 for aorta, and $j$7 for heart (Lambert et al., 2019). Those studies foreground organ delineation for radiotherapy planning, whereas CTseg foregrounds tissue probability estimation, diffeomorphic atlas alignment, and SPM-compatible morphometric outputs.

A common misunderstanding is therefore to treat CTseg as a generic CT segmentation package. The available evidence instead places it in a narrower but technically consequential niche: CT-specific unified segmentation for neuroimaging workflows, especially where spatial normalisation and volumetrics must remain compatible with SPM-style downstream analysis.