Deep Learning-Based Anatomical Segmentation of the Foot and Ankle: A Multi-View Radiograph Approach

We aimed to develop deep learning models that can identify and separate the shapes of 14 bones in the foot and ankle using multi-view radiographs and predict their masks.

We retrospectively collected 273 radiographs from 99 patients with anatomically normal feet, including anteroposterior (AP), oblique (OBL), and lateral (LAT) views, obtained between January 2020 and December 2021. In each view, 14 bones were segmented using AP and OBL radiographs and 6 bones using LAT radiographs. Ground truth masks were manually annotated by two radiology technologists and reviewed by an emergency medicine physician. Two deep learning models, a fully convolutional network (FCN) with ResNet-50 and DeepLabv3 with ResNet-50, were independently fine-tuned for semantic segmentation and evaluated using five-fold cross-validation.

In the AP and OBL views, both models attained mean intersection over union (mIoU) values ranging from 0.899 to 0.975 and from 0.875 to 0.978, respectively. In the LAT view, mIoU values varied from 0.926 to 0.976 for FCN-ResNet-50 and from 0.872 to 0.961 for DeepLabv3. FCN-ResNet-50 achieved slightly higher mIoU values than DeepLabv3, with statistically significant differences identified between the two models in the overall OBL view and across the 14 specific bones (p<0.05).

The FCN-ResNet-50 and DeepLabv3 models could be effective in automatically segmenting foot and ankle bone structures using multi-view radiographs.