- The paper introduces a unified probabilistic model that integrates tissue segmentation, spatial normalisation, and bias correction specifically tailored for CT imaging.
- The paper demonstrates significant improvements in segmentation accuracy and registration metrics, outperforming MRI-trained SPM pipelines in Dice scores and volumetric agreement.
- The paper validates CTseg’s robustness and efficiency through extensive experiments, enabling label-free deployment, accurate brain volumetrics, and effective sex classification in clinical scenarios.
Introduction
The clinical ubiquity of computed tomography (CT) in neuroimaging—particularly as a frontline diagnostic for acute cerebrovascular and traumatic events—creates an unmet need for robust quantitative analysis pipelines, as the dominant methods remain tailored to MRI. The distinct intensity characteristics of CT, especially diminished soft-tissue contrast and Hounsfield-based scaling, invalidate direct transference of MRI-based segmentation and normalisation workflows. "CTseg: A Tool for Brain CT Segmentation, Spatial Normalisation, and Volumetrics" (2605.05154) proposes and validates a generative-model-based solution specifically for non-contrast clinical CT, providing SPM-compatible outputs that unlock CT data for quantitative analysis pipelines traditionally reserved for MRI.
Methodological Framework
Multi-Brain Generative Model for CT
CTseg leverages the Multi-Brain framework—an extension of SPM's unified segmentation—with an integrated Gaussian mixture model (GMM) that jointly infers tissue segmentations (GM, WM, CSF, Bone, Soft tissue, Background), spatial priors (atlas), nonlinear deformation fields, and bias corrections in a single probabilistic generative model. Critically, atlas and intensity parameters are learned from multi-modal (CT and MRI) data, incorporating CT's unique intensity profiles and geometries while benefiting from MRI’s superior contrast to define precise tissue boundaries.
The training set spans large, public, and diverse neuroimaging data: CQ500 (CT), IXI, MICCAI 2012, and MRBrainS18 (MRI with expert annotations), ensuring the learned model captures generalisable intensity and spatial features across scanners and populations. The output format (native and warped tissue maps, deformation fields) directly mirrors SPM conventions, supporting downstream morphometric and lesion-mapping workflows without pipeline adaptation.
Atlas and Spatial Priors
CTseg offers both a data-driven groupwise-optimal atlas and SPM-aligned versions (1.5 mm and 1.0 mm isotropic resolutions), enabling analyses in MNI space and seamless integration with SPM-based toolchains. The model's spatial priors include expanded caudal coverage (upper cervical spine), facilitating consistent registration even in clinical CTs with large fields of view.
Unsupervised, Label-Free Deployment
A distinguishing feature is its unsupervised, label-free operation: CTseg requires no supervised training or retraining between sites, modalities, or protocols, providing rapid, robust volumetric and spatial normalisation outputs without GPU dependency.
Validation Strategy
Experimental Design
Validation utilises 59 paired, spatially aligned MR/CT scans from the SynthRAD2025 challenge, providing a rigorous evaluation against MRI-based SPM segmentation ("silver standard"). SPM's MRI-trained unified segmentation pipeline is also applied directly to CT as an out-of-domain baseline, enabling an informative comparison focused purely on the efficacy of CT-specific modelling.
Three evaluation axes are employed:
- Tissue segmentation accuracy: Dice, HD95, and ASSD for GM, WM, CSF versus MRI reference.
- Spatial normalisation: Sharpness and consistency of group-average normalised CT, voxelwise intensity CoV, and deformation field accuracy with respect to anatomical warping.
- Volumetric agreement: TBV, TIV, ICC, correlation, and Bland-Altman analysis versus MRI references.
- Sex classification from normalised tissue maps: Kernel-based GP classification as an indirect measure of morphometric fidelity in downstream applications.
External Benchmarking
CTseg’s practical impact is further documented by its adoption in over a dozen published clinical studies across stroke, neurodegeneration, radiomics, and multimodal PET integration, confirming real-world robustness and versatility.
Segmentation Accuracy
CTseg delivers statistically significant improvements over SPM-CT for all tissue classes:
- Dice (median, GM/WM/CSF): 0.601/0.721/0.466 (CTseg) vs. 0.531/0.659/0.367 (SPM-CT; all p < 0.001).
- ASSD (mm): Systematically lower for CTseg across all tissues, indicating more precise boundaries.
Notably, CSF segmentation—traditionally problematic for CT due to absent contrast—is improved by ~0.1 Dice by CTseg, underscoring the importance of CT-specific intensity modelling.
Spatial Normalisation
- Mean voxelwise intensity CoV (in brain): 0.306 (CTseg) vs. 0.402 (SPM-CT), reflecting higher-normalisation consistency.
- Dice for GM/WM after deforming native-space MRI segmentation: 0.667/0.808 (CTseg) vs. 0.633/0.764 (SPM-CT; p < 0.001).
These metrics demonstrate that CTseg's deformation fields achieve more accurate anatomical alignment, enabling high-fidelity population-based morphometry on clinical CT.
Brain Volumetrics
- TBV agreement (ICC): 0.829 (CTseg) vs. 0.650 (SPM-CT).
- TIV agreement (ICC): 0.821 (CTseg) vs. 0.798 (SPM-CT).
- Systematic negative bias for CTseg relative to the MRI reference observed for TBV and TIV, partially attributable to brain class definition differences and unavoidable limitations of CT contrast for CSF.
The authors warn against over-interpretation of closer mean volume agreement between SPM-CT and MRI, attributing such results to shared-method bias rather than true anatomical fidelity.
Predictive Application: Sex Classification
CTseg-normalised volumetric tissue maps enable kernel GP classification of sex with an AUC of 0.920, outperforming both SPM-MR (0.886) and SPM-CT (0.869). While the small, unbalanced cohort precludes robust significance analysis, these results support the preservation of sex-differentiating morphology by CTseg's registration model.
Median segmentation time is ∼2–3 minutes per subject on standard CPU hardware, with a 2× penalty over SPM, reflecting the computational load of advanced diffeomorphic registration. However, the pipeline remains viable for clinical and large-cohort research applications without GPU acceleration.
Practical and Theoretical Implications
CTseg presents a substantial advance in bridging the analytic gap between clinical CT and MRI-derived computational neuroimaging. Its generative, unsupervised framework provides:
- Robust domain adaptation without retraining or labelled data, enhancing translational potential across sites and scanner protocols.
- SPM compatibility, enabling reuse of decades of tool development for quantitative image analysis.
- Direct volumetric and morphometric analysis of vast retrospective clinical CT repositories, unlocking biomarkers (GBV, TIV, regional atrophy) for disease stratification, outcome prediction, and computational phenotyping.
Contrasting with supervised deep learning methods, CTseg offers interpretability, generalisability, and full-chain analysis (segmentation, deformation fields, volumetrics) in a single call. However, its performance—particularly in CSF segmentation—remains bounded by intrinsic limitations of CT imaging and the ceiling of generative model expressivity relative to high-capacity CNNs deployed in domain-specific, labelled settings.
The generative design supports direct extension to non-brain anatomies by retraining the model on appropriate CT targets—positioning CTseg as a generalisable approach to organ-level segmentation and normalisation in low-contrast modalities.
Future Directions
- Direct benchmarking against modern CNN-based and transformer-based segmentation networks trained on labelled CT data.
- Large-scale, multi-centre validation across diverse pathologies and demographics (paediatric, extensive brain pathology).
- Expansion to additional anatomical regions and integration with multimodal data (CT+PET, CT+MRI).
- Optimisation for accelerated CPU/GPU deployment to support real-time clinical scenarios.
Conclusion
CTseg establishes a rigorous, model-based approach to brain CT segmentation and normalisation that addresses the modality's long-standing analytic limitations. It outperforms MRI-trained SPM pipelines in both segmentation and spatial alignment, enables accurate brain volumetrics, and demonstrates practical value across diverse clinical applications. The method balances accuracy, interpretability, and generalisability, substantially enhancing the quantitative neuroimaging toolkit available to both researchers and clinicians working with routine CT data (2605.05154).
