X-DigiSkull: Digital Skull Imaging & Modeling

• X-DigiSkull is a digital imaging and modeling framework that converts CT and radiographic data into anatomically accurate 3D skull representations.
• It integrates multimodal workflows—CT segmentation, mesh repair, 3D printing, and cross-modal translation—for applications ranging from education to forensic identification.
• The ecosystem supports interdisciplinary tasks such as skull-to-face reconstruction and phylogenetically conditioned modeling, underpinning advanced diagnostic and investigative workflows.

Searching arXiv for the cited X-DigiSkull-related papers and neighboring work. X-DigiSkull is best understood, in the literature summarized here, as a skull-centered digital imaging and modeling ecosystem rather than a single canonical package. The name appears explicitly as a dataset for aligned synthetic and real skull radiographs in "MedShift: Implicit Conditional Transport for X-Ray Domain Adaptation" (Caetano et al., 29 Aug 2025). In adjacent work, the same label is used in technical blueprints that map skull-focused research into end-to-end workflows for CT/CBCT segmentation, mesh repair, 3D printing, skull-to-face reconstruction, cross-modal identification, craniofacial superimposition, motion registration, and phylogenetically conditioned skull generation. This suggests an "X-DigiSkull stack" (Editor's term): acquisition, structural representation, cross-modal translation, and task-specific inference.

1. Scope and research contexts

Within this broader stack, X-DigiSkull spans at least six research contexts. First, it covers educational and anatomical fabrication, where CT-derived skull meshes are repaired and manufactured as 1:1 printed teaching models (Shen et al., 2017). Second, it includes clinical CMF image analysis, where high-resolution CBCT/CT volumes are segmented and annotated with large landmark sets by coarse-to-fine CNN systems (Liu et al., 2021). Third, it includes radiographic domain adaptation, where aligned synthetic and real skull X-rays are used to benchmark unpaired translation under dose and appearance shifts (Caetano et al., 29 Aug 2025).

Fourth, X-DigiSkull includes skull-to-face inference and editing, ranging from anatomy-guided 3D facial reconstruction to skull-conditioned diffusion models and skeleton-consistent parametric generators (Liang et al., 2024). Fifth, it includes forensic identification and superimposition, where skull images, face images, and facial photographs are matched or co-registered through transform learning, graph matching, or evolutionary optimization (Singh et al., 2018). Sixth, it extends beyond human forensic workflows to comparative morphology, where skull generation is conditioned on phylogeny under extreme data scarcity (Lau et al., 28 Apr 2026).

A recurring misconception is to treat X-DigiSkull as a single benchmark or product. The cited work indicates a different structure: one explicit dataset name, plus a set of interoperable skull-centric pipelines whose common denominator is the conversion of skeletal anatomy into digital representations that support analysis, fabrication, or identity inference.

2. CT-derived skull modeling and physical fabrication

In its most literal anatomical form, X-DigiSkull begins with CT-derived skull reconstruction and mesh remediation. A practical workflow for the anatomy education setting uses a Siemens SOMATOM Force CT scanner in the axial plane, with slice width 1 mm, image production interval 1 mm, and helical pitch 0.9; the output is exported as DICOM (Shen et al., 2017). The digital pipeline proceeds through Mimics, 3ds Max, Geomagic Studio, Mudbox, and Meshmixer. In Mimics, thresholding uses a minimum value of 226 for Bone (CT), non-skull structures are removed, and thin or missing bony bridges are reconnected slice by slice. 3ds Max is used to cut and remove the skull cap. Geomagic then closes open boundaries, fills holes, remeshes, smooths, and decimates the surface; the reported interface state after improvement shows "Current Triangles: 204,446." Mudbox enhances anatomically important but under-resolved structures, including the supra-orbital foramen, infra-orbital foramen, anterior clinoid processes, and mastoid processes, and Meshmixer performs final printability inspection (Shen et al., 2017).

The printed artifact is produced on an Ultimaker2 FDM printer using white PLA filament of diameter 1.75 mm and layer thickness 0.1 mm. Reported timings are approximately 6 hours for digital repair and restoration, 28 hours for printing, and approximately 2.5 hours for hand painting with twelve different colors of propylene paint. Raw materials per skull are reported as approximately \$14, and the printer cost as approximately \$500 (Shen et al., 2017).

Validation in this setting is qualitative rather than metric-driven. The printed model is compared with a cadaveric skull in frontal, left, right, and anterior views, and is described as reproducing the same structures clearly at a 1:1 scale. The same study notes that CT accuracy is limited to approximately 1 mm and explicitly does not report dimensional deviation metrics such as Dice coefficient or Hausdorff distance. That absence matters: the workflow is validated as a practical educational surrogate, not as a quantitatively certified geometric standard (Shen et al., 2017).

3. Structural representations: voxels, landmarks, parametric skull-face models, and neural fields

At the clinical analysis layer, X-DigiSkull is instantiated as a segmentation-and-landmarking system. SkullEngine evaluates a multi-stage coarse-to-fine CNN framework on 170 CBCT/CT images, segmenting midface and mandible while detecting 175 landmarks distributed across bones, teeth, and soft tissues (Liu et al., 2021). The volumes are large and anisotropic, with median spacing $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$ and median matrix $512 \times 512 \times 418$. The architecture uses a scalable joint segmentation-and-detection stage on downsampled data, followed by three refinement models for global skull segmentation, thin facial bone recovery, and local tooth landmark detection. Reported test performance reaches DSC $88.5 \pm 6.9\%$ for midface and $93.5 \pm 3.4\%$ for mandible, with RMSE $3.03 \pm 1.96\ \mathrm{mm}$ for bony landmarks, $2.10 \pm 2.89\ \mathrm{mm}$ for tooth landmarks, and $3.34 \pm 3.20\ \mathrm{mm}$ for facial landmarks; inference completes within approximately 3 minutes per volume (Liu et al., 2021).

A different representation couples skull and face in a unified parametric model. SCULPTOR is trained on the LUCY dataset, which contains 144 CT scans of 72 subjects, each with pre- and post-orthognathic surgery scans, acquired at $0.48 \times 0.48 \times 1\ \mathrm{mm}^3$ voxels (Qiu et al., 2022). The model jointly represents skull, face geometry, and face appearance through shape, trait, pose, and expression factors, with anatomical jaw joints regressed from skull vertices. The dataset includes 29 skeletal landmarks and 15 facial landmarks. Reported results show skull fitting mean squared error of $1.77\ \mathrm{mm}$ for both pre- and post-surgery cases in the full model, and face reconstruction MSVE of $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$0 pre-surgery and $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$1 post-surgery for the configuration that combines shape and trait spaces (Qiu et al., 2022).

X-DigiSkull also includes neural implicit morphology. PhyloSDF operates on 100 micro-CT-scanned skulls across 24 species of Darwin's Finches and relatives, using a DeepSDF auto-decoder regularized by a Phylogenetic Consistency Loss and a Residual Conditional Flow Matching generator (Lau et al., 28 Apr 2026). The latent space is explicitly aligned with evolutionary distances, achieving Pearson $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$2, and the generative model produces 180 non-memorized meshes. Residual CFM reaches Chamfer Distance $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$3 and morphometric Fréchet distance $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$4, while standard flow matching collapses to $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$5–$0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$6 of real intra-species variation and denoising diffusion fails entirely at this data scale (Lau et al., 28 Apr 2026). In this context, X-DigiSkull is no longer only a human imaging pipeline; it becomes a general framework for skeletal morphology as a learnable latent object.

4. Radiographic datasets and synthetic-to-real domain adaptation

The most explicit named instance of X-DigiSkull is the dataset introduced by MedShift for X-ray domain adaptation (Caetano et al., 29 Aug 2025). It comprises aligned synthetic and real skull X-rays under varying dose settings, with geometric consistency rather than true pixel-wise pairing. The synthetic domains are Low and High dose; the real domains are Low, Normal, and Exposure, the last described as a Philips-exclusive mode providing enhanced image quality and detail. The dataset contains 18,220 images in total, split into 15,525 training and 2,695 test images. Synthetic images contribute 11,664 images and real images 6,556. All released images are cropped and resized to $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$7 pixels, while MedShift itself uses a latent-space encoder with input size $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$8 (Caetano et al., 29 Aug 2025).

The real images are acquired on a Philips Azurion Image Guided Therapy system from common neuro-interventional working positions. The synthetic images are generated with the Mentice VIST® G7 simulator, which models per-voxel attenuation and then approximates scatter, focal-spot and detector blur, grid and heel effects, detector response, quantum and read noise, and final post-processing. Real and synthetic coordinate systems are aligned so that synthetic viewpoints approximate phantom acquisitions, but the dataset remains unpaired in the strict supervised sense (Caetano et al., 29 Aug 2025).

The benchmark task emphasized in the paper is Synthetic High $0.39 \times 0.39 \times 1.0\ \mathrm{mm}^3$9 Real Normal translation. MedShift evaluates realism with CFID, Coverage, and CMMD, and structure preservation with LPIPS and SSIM. At $512 \times 512 \times 418$0, MedShift reports CFID $512 \times 512 \times 418$1, Coverage $512 \times 512 \times 418$2, CMMD $512 \times 512 \times 418$3, LPIPS $512 \times 512 \times 418$4, and SSIM $512 \times 512 \times 418$5; at $512 \times 512 \times 418$6, it reports CFID $512 \times 512 \times 418$7, Coverage $512 \times 512 \times 418$8, CMMD $512 \times 512 \times 418$9, LPIPS $88.5 \pm 6.9\%$0, and SSIM $88.5 \pm 6.9\%$1, making explicit the structure-style trade-off (Caetano et al., 29 Aug 2025). The same study reports inference latency of $88.5 \pm 6.9\%$2–$88.5 \pm 6.9\%$3 on RTX 5090 and model size $88.5 \pm 6.9\%$4. It also notes important limitations: a single skull phantom, one simulated head model, approximate geometric alignment rather than true pairs, and the absence of segmentation or diagnostic labels.

5. Cross-modal identification and skull-conditioned facial synthesis

A major branch of X-DigiSkull concerns the mapping from skull evidence to facial identity. Early automated skull-face matching is formulated as a heterogeneous identification problem on the IdentifyMe dataset, the first public skull-face image pair database introduced for this purpose (Nagpal et al., 2017). IdentifyMe contains 464 skull images, including 35 labeled skull-face pairs and 429 unlabeled skull images intended for unsupervised training. Under five-fold cross-validation, Semi-Supervised Transform Learning with HOG features reports rank-1 identification accuracy of $88.5 \pm 6.9\%$5 on Protocol-1 and $88.5 \pm 6.9\%$6 on Protocol-2, where the latter uses a gallery of 1,000 face images (Nagpal et al., 2017). A subsequent Shared Transform Model learns a single transform for both skull and face images and explicitly reduces intra-class distance between paired representations; with HOG features it reports rank-1 accuracy of $88.5 \pm 6.9\%$7 on Protocol-1 and $88.5 \pm 6.9\%$8 on Protocol-2, together with improved robustness across folds (Singh et al., 2018). By contrast, the earlier "3D Skull Recognition Using 3D Matching Technique" is primarily a conceptual note: it asserts fast processing and optimization of false positives and negatives, but does not specify acquisition sensors, registration mathematics, datasets, or quantitative performance (Alanazi et al., 2010).

More recent work shifts from direct retrieval toward generative or graph-structured cross-modal inference. Skull-to-Face reconstructs a textured 3D face from a digital skull by combining Stable Diffusion XL, DECA, FLAME, and tissue-thickness priors learned over 78 landmarks; the core anatomical relation is $88.5 \pm 6.9\%$9 (Liang et al., 2024). On Skull100, the method reports normalized mean error $93.5 \pm 3.4\%$0 for average tissue settings, approximately $93.5 \pm 3.4\%$1, and $93.5 \pm 3.4\%$2 for the best among thin, normal, and fat variants (Liang et al., 2024). Cranio-Diff instead uses 2D X-ray skulls as ControlNet guidance with biometric text conditioning for age, gender, orientation, and BMI. On a 120-subject skull-face dataset expanded to 4,320 paired samples, it reports FID $93.5 \pm 3.4\%$3, SSIM $93.5 \pm 3.4\%$4, LPIPS $93.5 \pm 3.4\%$5, PSNR $93.5 \pm 3.4\%$6, and ArcFace score $93.5 \pm 3.4\%$7, outperforming CycleGAN, Pix2Pix, and ICCR-Diff on the reported metrics (Prasad et al., 8 Jun 2026).

Graph-based retrieval provides a third path. Cranio-ID automatically detects landmarks with YOLO-Pose on skull X-rays and face images, converts them into graphs, applies cross-attention, and solves cross-modal correspondence with entropic optimal transport (Prasad et al., 18 Nov 2025). On the S2F benchmark, the ViT backbone with both OT and cross-attention reports $93.5 \pm 3.4\%$8, $93.5 \pm 3.4\%$9, $3.03 \pm 1.96\ \mathrm{mm}$0, $3.03 \pm 1.96\ \mathrm{mm}$1, with $3.03 \pm 1.96\ \mathrm{mm}$2 (Prasad et al., 18 Nov 2025). These results do not remove forensic uncertainty. Both the reconstruction and retrieval literature repeatedly imply that outputs are best used as candidate-generation or investigative aids rather than identity proofs.

6. Registration, superimposition, and principal limitations

Forensic and biomechanical registration form another major branch of X-DigiSkull. In craniofacial superimposition, the key task is Skull-Face Overlay: aligning a 3D skull model to a 2D ante-mortem facial photograph while accounting for uncertainty in soft-tissue depth. Lilium models that uncertainty with 3D cones attached to cranial landmarks and optimizes the cone parameters by Differential Evolution under landmark, camera, boundary, and region-parallelism constraints (Martínez-Moreno et al., 26 Feb 2026). On synthetic experiments derived from 17 CT scans and 17,340 SFO instances per run, Lilium substantially reduces anatomically implausible overlays. Worst-case implausible overlays fall from $3.03 \pm 1.96\ \mathrm{mm}$3–$3.03 \pm 1.96\ \mathrm{mm}$4 for POSEST-SFO to $3.03 \pm 1.96\ \mathrm{mm}$5–$3.03 \pm 1.96\ \mathrm{mm}$6 with boundary and parallelism constraints, while the fraction of skull pixels outside the facial mask drops to $3.03 \pm 1.96\ \mathrm{mm}$7–$3.03 \pm 1.96\ \mathrm{mm}$8 (Martínez-Moreno et al., 26 Feb 2026). The improvement is computationally expensive: deterministic POSEST-SFO runs in approximately $3.03 \pm 1.96\ \mathrm{mm}$9–$2.10 \pm 2.89\ \mathrm{mm}$0 per SFO, whereas Lilium requires approximately $2.10 \pm 2.89\ \mathrm{mm}$1–$2.10 \pm 2.89\ \mathrm{mm}$2 when containment and parallelism are enforced (Martínez-Moreno et al., 26 Feb 2026).

In motion-tracking applications, X-DigiSkull includes 3D-to-2D registration of a skull model to dual fluoroscopic X-rays. A transfer-learning strategy creates 9,751 DRR-landmark pairs from a single CT, trains a ResNet landmark regressor and a CycleGAN for DRR-X-ray style translation, and then optimizes skull pose from detected landmarks in a non-orthogonal dual-fluoro setup (Zhou et al., 2021). In walking, reported angular and position errors are $2.10 \pm 2.89\ \mathrm{mm}$3 and $2.10 \pm 2.89\ \mathrm{mm}$4; performance degrades in functional neck activity, where partial skull visibility becomes severe (Zhou et al., 2021). This limitation is characteristic of the broader X-DigiSkull literature: performance is often bounded by sparse labels, approximate pairing, thin or occluded structures, scanner artifacts, and the fact that skull information alone underdetermines soft-tissue appearance.

Across all subfields, validation remains heterogeneous. Educational printing emphasizes qualitative comparison to cadaveric references (Shen et al., 2017). Retrieval systems report rank-based metrics on small or moderate datasets (Nagpal et al., 2017). Generative systems emphasize FID-, LPIPS-, SSIM-, and identity-style metrics (Prasad et al., 8 Jun 2026). Superimposition emphasizes back-projection error, rank, and anatomical plausibility (Martínez-Moreno et al., 26 Feb 2026). This diversity suggests that X-DigiSkull is methodologically unified by skull-centered computation, but not yet standardized by a single evaluation protocol.