mikrooo1595/ct-skull-quad-retopo-pack

CT-Derived Human Skull Mesh Pack: Quad-Retopologized, Quality-Measured (A0102–A0105)

Author: Iqbal Muhammad Yasin (ORCID: 0009-0007-2277-6917) DOI: 10.5281/zenodo.21533007 License: CC-BY-4.0 Source data: MUG500+ skull CT dataset (Li et al., 2021) — CC-BY-4.0, DOI:10.6084/m9.figshare.9616319

Summary

This deposit provides documentation, methodology, and measured quality reports for four human skull meshes derived from clinical CT scans (MUG500+ dataset), retopologized to 100% clean quad topology using Blender QuadriFlow. Each specimen is processed under an identical, reproducible pipeline and measured against its source CT surface — not visually estimated.

The full-resolution mesh files (OBJ/FBX/GLB/STL) are available for purchase via https://www.fab.com/listings/5570c50f-d58a-4553-ac21-b312a181e50c. This Zenodo deposit exists to make the methodology, quality metrics, and citation record permanently and freely accessible, independent of any commercial marketplace.

Methodology

Pipeline: MUG500+ skull CT (segmentation mask) → canonical frame alignment → adaptive neck cut (v2) → field-blur (σ=0.8) → QuadriFlow retopology with a micro-cleanup pre-pass (bersih_mikro: dissolve degenerate geometry + strip non-manifold vertices + recalculate normals) → island removal.

Tools used: Blender QuadriFlow (BSD/MIT), numpy/scipy/scikit-image (BSD). No GPL-licensed tooling was used in the pipeline.

Known issue and fix (documented for reproducibility): the initial smoothing/wrap step introduced zero-length edges and non-manifold vertices that were not caught by a zero-area face check alone, causing QuadriFlow to silently reject the mesh and fall back to a triangulated export. This was resolved by adding the bersih_mikro cleanup pass (degenerate-edge dissolve + non-manifold vertex removal + normal recalculation) before QuadriFlow — quad yield went from 0% to 100% across all four specimens after this fix.

Quality cards (measured, not estimated)

Skull	Quad faces	100% quad	2-manifold	Fidelity — Hausdorff p95 (mm, ≤2.0 threshold)	Step-angle p95
A0102	74,558	✅	✅	1.44	62.8°
A0103	71,779	✅	✅	1.45	66.1°
A0104	72,626	✅	✅	1.50	61.8°
A0105	75,288	✅	✅	1.36	57.9°
A0101 (σ0.8 reference)	76,022	✅	✅	1.37	50.6°

All four specimens fall within the same fidelity class as the reference specimen (~1.4mm Hausdorff p95). 100% quad and 2-manifold: pass on all.

Inter-specimen variation (step-angle 57.9°–66.1° vs. the A0101 reference of 50.6°) is quantified and preserved as genuine anatomical variation between specimens, not cleanup noise or inconsistency artifacts — this makes the set valid as matched comparison/training material where uniform processing matters.

Honest limitations (stated explicitly)

• Jaw/teeth regions still reflect raw CT complexity in some specimens (e.g., metallic dental artifacts); a dedicated dental-finishing pass is planned as a separate deliverable.
• The neck is intentionally open (a display surface, not a solidified volume) across all specimens, consistent with the A0101 reference.
• STL exports are not fully watertight; closing the neck opening in a slicer is required for solid 3D printing.
• This is not a medical diagnostic tool. Not validated for clinical use.

Reproducibility

Fixed random seed (seed=42); output verified byte-identical across repeated runs. Full processing scripts (retopo_quadriflow.py with the bersih_mikro pre-pass, buang_pulau.py) referenced in the processing log; available on request for verification purposes.

Citation

If you use this methodology or quality-measurement approach in your work, please cite:

@dataset{yasin2026skullpack,
  author       = {Yasin, Iqbal Muhammad},
  title        = {CT-Derived Human Skull Mesh Pack: Quad-Retopologized,
                   Quality-Measured (A0102-A0105)},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.21533007},
  note         = {Source data: MUG500+ (Li et al., 2021, CC-BY-4.0,
                   DOI:10.6084/m9.figshare.9616319)}
}

Please also credit the original data source per its license terms:

Li, J. et al. (2021). MUG500+: Database of 500 high-resolution healthy
human skulls and 29 craniotomy skulls and implants. Data in Brief.
DOI: 10.6084/m9.figshare.9616319. CC-BY-4.0.

Contact

For the full-resolution mesh files, commercial licensing, or dataset licensing for ML training use, see: (https://www.fab.com/listings/5570c50f-d58a-4553-ac21-b312a181e50c) or contact [email protected].

github : https://github.com/mikrooo1595/ct-skull-quad-quality-reports?tab=readme-ov-file

Available for purchase via Fab (https://www.fab.com/listings/5570c50f-d58a-4553-ac21-b312a181e50c) or CGTrader (https://www.cgtrader.com/3d-models/character/human-anatomy/human-skull-3d-model-anatomically-accurate).