Automated International Hip Dysplasia Institute Classification Using Deep Convolutional Neural Networks

Developmental dysplasia of the hip (DDH) is hip instability or dislocation due to developmental aberrations of hip anatomy. Hip radiographs are commonly used to diagnose and classify DDH using methods such as the IHDI classification. This classification helps clinicians in predicting the natural history of DDH as well as the response to treatment. Manual measurement can be time-consuming and prone to intra- and inter-rater variability, limiting the ability to accurately compare outcomes across large datasets. The purpose of this study was to develop and test an automated method to determine the IHDI classification on hip radiographs. Utilizing radiographs collected as part of a global, prospective registry of infants and children diagnosed with DDH, we used a machine learning approach based on convolutional neural networks (CNNs) to segment pediatric hip radiographs and extract the anatomy of interest. A custom Matlab code was then applied to extract anatomic landmarks and to determine the IHDI classification (Figure 1). Automated IHDI classification was compared to expert manual measurement by orthopaedic surgeons. Figure 1. Overall CNN output with overlayed landmarks used to determine IHDI classification. A total of 232 radiographs were manually measured by two orthopaedic surgeons (464 hips). Of these, 221 were deemed appropriate and successfully processed by all three CNNs. We compared the manual to automated processes in terms of differentiating dysplastic and non-dysplastic hips and IHDI classification. We demonstrated an accuracy of 75.8%, a precision of 69.0% and a sensitivity of 83.8% in detecting DDH. With respect to classifying DDH using the IHDI classification system, we show a true positive rate of 83.3%, 27.5%, 59.3%, 93.6% for IHDI grades 1–4, respectively. The CNN-based workflow presented in this work is shown to be capable of determining the IHDI classification of pediatric hip radiographs. We show accuracy and precision inferior to orthopedic surgeons but which can be improved as the registry dataset continues to grow. We foresee such applications to have the potential to facilitate larger research projects that would otherwise be restricted by the availability of expertise and time resources, in addition to standardizing measurements across diverse datasets. The IHDI classification is widely used both for diagnosing DDH, as well as for tracking progression, both clinically and in research. Therefore, an automated, fast and replicable method for classification has the potential to help in patient care in remote settings as well as in research projects involving large datasets. For any figures or tables, please contact the authors directly.