Recent advances in 3D printing with biocompatible materials have opened the door to patient-specific medical implants (see our 3D-Printed Implants). However, designing them is still a time-consuming and largely manual process, typically performed by trained experts.

To streamline this, we’ve developed a novel AI-based approach that automates implant design. At the heart of our system is a powerful type of machine learning model known as a Denoising Diffusion Probabilistic Model (DDPM), originally developed for image generation. We've adapted this technology to work with 3D shapes, enabling it to learn how to reconstruct missing parts of a patient’s skull from CT scans.

Here's how it works: The AI first analyzes a scan of a skull with a defect. It then predicts what the complete, healthy skull should look like, based on patterns learned from many previous examples. This is done by gradually transforming a cloud of random points into a detailed 3D structure — a bit like sculpting from digital noise. The missing bone structure is inferred by comparing the damaged skull to the predicted complete version, and the difference becomes the implant design.

Our method can generate implants for a wide range of defect types, from small gaps to complex, multi-region damage like fronto-orbital or bilateral defects. Despite this complexity, the implant generation typically completes in just 20 minutes — a fraction of the time compared to traditional manual design.

This breakthrough paves the way for faster, more accessible, and highly precise implant creation, bringing personalized medicine one step closer to widespread clinical practice.