Plausible but unimplantable: audit finds shape completion foundation models fail clinical cranioplasty constraints

An empirical audit of 3D shape completion foundation models reveals that despite visual plausibility, zero-shot reconstructions penetrate patient bone and achieve a 0% clinical implantability rate.

Primary source: arXiv:2610.30035: https://arxiv.org/abs/2610.30035

What’s new: Designing patient-specific cranial implants after craniectomy currently requires manual computer-aided design (CAD) by surgical specialists, a slow and expensive process that restricts clinical accessibility. While 3D shape completion foundation models claim zero-shot geometric reconstruction capabilities, their clinical safety has remained largely untested. In Plausible but Unimplantable: A Trustworthiness Audit of Shape Completion Foundation Models for Cranial Defect Reconstruction (arXiv:2610.30035, 21 Oct 2026), Marek Wodzinski, Kamil Kwarciak, Daria Hemmerling, and Mateusz Daniol evaluate three public pretrained shape completion models, including two foundation models (MedShapeNet-FM and MGPC) alongside AdaPoinTr, across 239 cranial defects (210 synthetic defects from SkullBreak/SkullFix and 29 real-world clinical craniectomy cases from MUG500+).

The researchers benchmarked zero-shot completions against a strict physical implantability criterion requiring collision freedom within a 1 mm tolerance of the patient’s remaining bone and a positive volumetric addition over a trivial do-nothing baseline (simply returning the skull as-is). Strikingly, across 2,868 model configurations, every audited foundation model achieved a 0% implantability rate (0/239 skulls). On synthetic defects, none of the models outperformed doing nothing, and on clinical cases, predictions penetrated patient bone by median depths of 4.4 mm to 6.6 mm. The authors show that surface-level point cloud representations (at standard 6,144-point budgets) fail to resolve plate thickness, meaning visual realism completely masks physical incompatibility. In contrast, a task-specific 3D volumetric convolutional baseline achieved clinical implantability on 238 of 239 skulls.

Why it matters: Generative foundation models for 3D anatomy often produce visually convincing reconstructions that score well on standard computer vision metrics like Chamfer distance. However, in surgical reconstruction, an implant that intersects existing cranial bone or lacks uniform structural thickness cannot be seated in an operating room without risking brain compression or mechanical failure. This audit underscores that clinical deployment requires physics-informed constraints, volumetric formulations, and baseline comparisons against trivial solutions rather than unconstrained point cloud generation.

Caveats: The audit evaluated foundation models in zero-shot configurations without task-specific fine-tuning on cranial defect pairs. While dedicated 3D volumetric segmentation models readily passed implantability checks, the compute and annotation costs required to train specialized networks for every anatomical defect type remain high.