Deep learning models segment 14 foot and ankle bones on multi-view radiographs

An automated segmentation model (FCN-ResNet-50) identified 14 foot and ankle bones on AP and oblique radiographs with a mean intersection over union (mIoU) of 0.899–0.975 and 6 bones on lateral views with mIoU of 0.926–0.976, slightly outperforming DeepLabv3.

• The study trained and tested two deep learning models (FCN-ResNet-50 and DeepLabv3) on 273 radiographs from 99 patients with normal foot anatomy.
• Ground truth masks were manually annotated by two radiology technologists and reviewed by an emergency medicine physician.
• The study was retrospective and used five-fold cross-validation without an external test set.

Automated summary

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