Schatzker classification deep learning benchmark controlled with inexpensive protocols reveals inflated signal from non-fracture features

A ResNet-50 on 421 AP knee radiographs with expert labels reached 0.345 balanced accuracy for six-class Schatzker classification, only 0.168 above a non-anatomical baseline (95% CI 0.106–0.230, p=0.002). Tibia-only crops performed better than whole radiographs, and the mask alon…

• On fracture versus no-classifiable-fracture discrimination, the network achieved 0.833 balanced accuracy, not significantly better than a model without access to pixel data (0.814, p=0.264).
• Recall was graded: 0.72 for Schatzker type VI but only 0.11 for type V and 0.04 for type IV (chance 0.167).
• Controls including complement ablation and augmentation audit showed that much of the performance attributed to fracture detection stemmed from anatomical context and data-set biases.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.