Tailoring Patient-Specific Cranial Implants for Bone Reconstruction via End-to-End Deep Learning Image-to-Print Approach

Timely, patient-specific cranial implants are critical for restoring skull integrity after trauma, accidents or surgery, yet off-the-shelf devices arrive in coarse, fixed size increments that often fail to match irregular defects. Surgeons are therefore forced to open and try multiple implants intra-operatively, adding infection risk, operative time, and likelihood of early implant failure. This paper presents an end-to-end workflow that predicts “print-ready” cranial implants directly from defective CT volumes and validates them through physical prototyping. The pipeline couples a five-stage volumetric preprocessing routine with a systematic exploration of neural architectures. A baseline and two deeper 3-D U-Nets are benchmarked alongside the proposed version 3 model, a channel-rebalanced U-Net that shifts capacity from shallow texture filters to boundary-aware decoding paths. Two public datasets, SciData and SkullFix, were merged, and split for training the model. On 40 unseen skulls, the proposed version 3 achieved a mean Dice of 0.901, Boundary Dice of 0.908, and HD95 of 1.52 mm, which is surpassing all alternatives and reducing the number of intra-operative trials predicted by simulation from 3-4 to zero. Bench-top tests confirmed that implants printed directly from the network’s STL output seated flush on three morphologically diverse phantoms without CAD edits or sanding, demonstrating true “push-button” manufacturability. These results show that a carefully optimized architecture, paired with robust pre-processing, can supply size-perfect implants using modest data and commodity hardware.