Abstract
Research Type:
Level 3 - Retrospective cohort study, Case-control study, Meta-analysis of Level 3 studies
Introduction/Purpose:
Accurate assessment of foot alignment on radiographs relies on precise identification of key bony landmarks. Manual annotation is time-consuming and subject to inter- and intra-observer variability. Automated landmark detection offers a promising solution by enhancing efficiency and reproducibility. In this study, we aimed to develop an automated method for detecting critical bony landmarks on foot radiographs.
Methods:
A retrospective study was conducted using 1,207 foot radiographs of individuals without a history of fractures or severe deformity. Three orthopedic researchers manually annotated bony landmarks in the hindfoot and forefoot using CVAT platform (CVAT Co, Palo Alto, CA). Anteroposterior radiographs were used to annotate four landmarks on the 1st proximal phalanx, 1st, and 2nd metatarsals, while lateral radiographs included seven calcaneus landmarks placed on key anatomical regions. The UNet2D model, a deep learning-based architecture designed for robust landmark detection, was employed for automatic landmark identification. The dataset was split 85:15 for training and validation. Model performance was assessed using the Mean Radial Error (MRE) ± Standard Deviation, quantifying the average deviation of predicted landmarks from the ground truth. Additionally, the Success Detection Rate (SDR) was used to measure accuracy, defined as the percentage of detected landmarks that fall within a 5 mm error threshold relative to ground truth annotations.
Results:
Forefoot landmark detection yielded an MRE of 2 ± 2.05, with an overall SDR of 94.0% at 5 mm. The 1st metatarsal bone had the lowest SDR at 89% (Figure 1). For the calcaneus, the model achieved an MRE of 3.76 ± 4.21 and an overall SDR of 79%, with the posterior facet landmark exhibiting the lowest SDR at 51% (Figure 1).
Conclusion:
This study demonstrates the feasibility of automated bony landmark detection on foot radiographs using deep learning. The model achieved high accuracy in detecting hindfoot and forefoot landmarks, providing a reliable foundation for further radiographic analysis and measurements. These detected landmarks can be utilized in post-processing to perform real-time 2D angular or linear measurements, enabling automated assessment of foot alignment. Our results lay the foundation for automated intraoperative measurements, which leads to reduced radiation exposure through fewer X-ray shots and supports the decision-making processes.
Figure 1. Bony landmark detection in (a) calcaneus bone, (b) 1st, and 2nd metatarsal bones along with the 1st proximal phalanx. The green dots indicate the ground truth, while the red dots represent the model’s predictions. The white arrow shows the posterior facet landmark which had the least SDR value.