Vision Transformer cracks ceramic implant fracture diagnosis with 90.7% accuracy

A team led by Julian Schmid developed an interpretable vision-transformer (ViT) workflow to automate fracture-cause classification in ceramic hip and knee implants made from BIOLOX delta (CeramTec GmbH). They curated a dataset of 8,493 scanning electron microscopy (SEM) images spanning magnifications from 50x to 10,000x, sourced from five years of in-production burst and proof tests. The images were annotated into three defect categories along the manufacturing chain: green body, hard machining, and material defects. Despite severe class imbalance, a fine-tuned ViT achieved 90.7% accuracy and a macro-F1 of 0.888 in stratified five-fold cross-validation. A rigorous two-stage perceptual-hash/SSIM leakage audit confirmed negligible specimen overlap.

The most striking finding: performance at low magnification (50x) was comparable to that at high magnification (1k–10kx). This suggests that macro-scale features — mirror geometry and hackle line fields — already encode sufficient diagnostic signal for fracture origin identification. Grad-CAM attributions consistently localized on canonical fractographic cues such as mirrors, hackles, pores, and machining marks, aligning with established human expert criteria. The work positions interpretable ViTs as a complementary tool for ceramic-implant quality assurance, enabling low-magnification pre-screening that reduces reliance on time-intensive high-magnification inspection and speeds up the QC pipeline.