Cross-Species Nano-CT Segmentation

Conventional micro-CT can resolve enamel rods, but only at the bulk scale — the nanoscale geometry that actually controls fracture toughness (rod diameter, packing, circularity) sits below most reconstruction noise floors. This paper shows that a paired self-supervised denoising + U-Net segmentation approach can recover that geometry across three species — human, African lion, African wild dog — and turn it into per-rod metrics suitable for comparative analysis and downstream biomimetic design.

• Species: human, African lion (Panthera leo), African wild dog (Lycaon pictus)
• Imaging: synchrotron nano-CT
• Output: per-rod diameter, circularity, spacing, orientation
• Downstream: STL meshes for manufacturability screening

Noise2Inverse is used as a self-supervised denoiser: the reconstruction is split into projection sub-sets, each sub-set is independently reconstructed, and a network is trained to map noisy reconstructions to their independent counterparts. No clean ground truth is needed — the noise model itself supplies the supervision.

A 2.5D U-Net is then trained on a small set of manually-annotated transverse slices to segment individual rods. From the segmentation we extract rod-level descriptors (diameter, axial profile, circularity, neighbor spacing) and aggregate them into species-level distributions for direct comparison.

Published in Acta Biomaterialia as corresponding author. The pipeline now feeds into the biomimetic design work in Project 01, where the cross-species rod geometries inform the parameter space for 3D-printed lattice prototypes.