MRI-guided physics-constrained deep generative modelling for personalized osteochondral scaffold design

Abstract Osteochondral defects affecting the bone-cartilage interface present a persistent clinical challenge because current scaffold-based repair methods rely on geometrically generic, uniformly porous implants that do not adapt to patient anatomy, defect morphology, or the zonal mechanical gradient of native osteochondral tissue. Existing Deep Learning (DL) methods either produce anatomy labels without generating scaffold designs (SD) or generate scaffold geometries (SGs) from abstract scalar property inputs without incorporating patient imaging, leaving no computational pathway from clinical MRI to a personalised, biomechanically informed scaffold output. MedSca3D is proposed as an anatomy-conditioned physics-constrained latent diffusion model for personalized osteochondral scaffold design , transforming routine knee MRI volumes into printable, physics-constrained, personalized osteochondral SD through four sequential computational stages. Three spatially structured conditioning fields are derived from OAI-3D-DESS knee MRI: a synthetic defect void mask delineating the scaffold target region, a normalised subchondral support field serving as an MRI-derived structural conditioning proxy, and a thickness-normalized cartilage depth field encoding the native osteochondral porosity gradient. A VQ-VAE compresses dual-channel scaffold volumes into a discrete latent space, within which a conditional latent diffusion model generates scaffold geometry (SG) and gradient porosity jointly, conditioned on the three anatomical fields through cross-attention injection. A frozen neural finite element analysis surrogate enforces von Mises stress compatibility as a differentiable training constraint without requiring finite element simulation at inference time. Evaluated on OAI-derived synthetic osteochondral defect test cases, MedSca3D achieves a Defect Volume Overlap of 0.934, Porosity Root Mean Square Error of 0.051, Gradient Profile Correlation of 0.923, Young’s Modulus Relative Error of 4.37%, and Stress Concentration Factor of 3.17. Against the procedural TPMS baseline, MedSca3D records lower leakage, higher defect conformity, lower porosity error, and lower stress concentration, while Rule-TPMS achieves a lower modulus error of 3.91%. Compared with the deterministic regression baseline, MedSca3D performs better across all reported metrics. The complete pipeline delivers a printable STL-SD in 43 seconds per case .