Abstract
Category:
Hindfoot; Trauma
Introduction/Purpose:
Fractures of the calcaneus are the most frequent form of fractures among the tarsal bones. The treatment plan depends on various fracture characteristics, particularly the extension of the fracture line into the subtalar joint surfaces. While X-rays are the primary diagnostic tool, Computed Tomography (CT) scans are often necessary as they provide a more detailed description of these fractures, especially if the articular surfaces are involved. Although informative, CT scans expose patients to higher doses of ionizing radiation, and they might not be readily available in certain settings. This study aims to develop a deep learning model that will first detect calcaneal fractures and then classify the type of fracture based on its extension into the subtalar joint surfaces.
Methods:
A retrospective case-control study was conducted. Patients aged >18 years who sustained a calcaneal fracture were included in the case group, along with the same number of healthy adults for the control group. Various X-ray views of the foot including anterior, posterior, lateral, along with the axial and lateral calcaneal views were obtained. Two experienced orthopaedic researchers screened the dataset and confirmed the type of fracture. Two models were developed using an Inception V3 architecture and a split ratio of 60:20:20 for training, validation, and testing purposes. The first model was designed for fracture detection, and the second one was designed for detecting intra-articular fracture extension, if present. Conventional model metrics including sensitivity, specificity, Youden index, area under the receiver operating characteristic curve (AUC), and F1 score were reported for each model. A p-value below 0.05 was statistically significant. Baseline values are presented in median and interquartile ranges (IQR).
Results:
A total of 1102 individuals (n=551 for each group) with a female-to-male ratio of 1.6:1 were included. Baseline comparisons showed a significant difference in the age of the groups [cases: 55 (IQR 42, 67) years; controls: 62 (IQR 51, 72 years), p-value < 0.001]. The first model for fracture detection demonstrated a high performance with an AUC of 0.99 and a Youden index of 0.96. The intra-articular fracture detection model showed a slightly lower performance with an AUC of 0.97 and a Youden index of 0.88 (Table 1, Figure 1).
Conclusion:
Our study showed promising results with the utility of deep learning methods for detecting calcaneal fractures and classifying intra- vs extra-articular fractures. This algorithm may help facilitate the diagnosis and treatment of patients with suspected calcaneal fractures, serving as a decision-support tool in settings where CT scanning is not available.
