Shape completion models fail to produce usable skull implants
Plausible but Unimplantable: A Trustworthiness Audit of Shape Completion Foundation Models for Cranial Defect Reconstruction
Computer Vision and Pattern Recognition
Summary
People who lose part of their skull need implants that fit exactly, but these implants are usually designed by hand, which takes a lot of time and effort. The authors tested several computer models that try to complete the missing skull shape automatically. They found that none of these pre-trained models produced implants that could actually be used in real surgeries, because the predicted shapes either overlapped with existing bone or didn’t add enough material. This means current AI models for this task are not reliable enough for medical use yet.
What this means in practice
- •For medical device designers: Understand the limitations of current shape completion AI when automating cranial implant designs to avoid unimplantable outcomes.
- •For 3d printing specialists: Improve quality checks by knowing that today’s shape completion models may produce unusable implant shapes even if visually plausible.
Authors
Marek Wodzinski, Kamil Kwarciak, Daria Hemmerling, Mateusz Daniol
Abstract
A patient who has lost part of the skull vault needs an implant that fits the defect exactly, and such implants are still drawn by hand in CAD by specialists, which makes each case slow and expensive. The task can be considered as a shape completion problem, so pretrained completion models, and now the first shape completion foundation models, look like a way to remove that manual step. We audited three publicly released pretrained shape completion models, two of them foundation models, applied zero-shot to 239 cranial defects, some carved from public skull datasets and some real post-craniectomy cases. Each output was compared with simply returning the patient's own remaining bone, and tested for whether the predicted material could be placed without intersecting the bone that is still there. On the 210 carved defects none of the three improved on doing nothing, and on the 29 real cases only one did so beyond chance. Yet not one output, from any model in any configuration, was implantable: none managed both to stay clear of the bone the patient still has and to place more material than doing nothing, where a task-specific volumetric model, trained for this task and selected on these cases, was implantable on all but one. The output that most convincingly resembles a skull is no more placeable than the rest, and at these point densities implant thickness cannot be read off the emitted points at all, which is a property of the output representation rather than of any one model. On this evidence the shape completion foundation models we could obtain cannot be trusted and, used as released, are not usable in clinical practice. Substantial research towards trustworthy foundation models for shape completion is still needed.