2023 ISAKOS Biennial Congress ePoster
A Deep Learning Approach Using an Ensemble Model to Auto-Create an Image-Based Hip Fracture Registry
Jacobien Oosterhoff, MD NETHERLANDS
Soomin Jeon, PhD, Boston UNITED STATES
Bardiya Akhbari, PhD, Boston UNITED STATES
David Shin, BSc, Boston UNITED STATES
Daniel Tobert, MD, Boston UNITED STATES
Synho Do, PhD, Boston UNITED STATES
Soheil Ashkani-Esfahani, MD, Boston, Massachusetts UNITED STATES
Hamid Ghaednia, PhD, Boston, MA UNITED STATES
Joseph Hasbrouck Schwab, MD, Boston, Ma UNITED STATES
Massachusetts General Hospital, Harvard Medical School, Boston, Massachusetts, UNITED STATES
FDA Status Not Applicable
Summary
A Deep Learning Approach to Auto-Create an Image-Based Hip Fracture Registry
ePosters will be available shortly before Congress
Abstract
Background
With more than 300,000 patients per year in the US alone, hip fractures are one of the most common fractures occurring in the elderly, and the incidence is predicted to rise to 6 million cases annually worldwide by 2050. Many fracture registries have been established, serving as tools for both quality surveillance and evaluating patient outcomes. Most registries are based on billing and procedural codes, which may underreport fracture cases. As deep learning (DL) has shown to be successful for interpreting radiography and assisting in fracture detection; we propose to conduct a DL-based approach intended to assist in creating a fracture registry, specifically for the hip fracture population.
Methods
Conventional radiographs (n=18,834) from 2,919 patients from the Massachusetts General Brigham hospitals were extracted (images designated as hip radiographs within the medical record). We designed a cascade model consisting of three submodules for image view classification (M1), postoperative implant detection (M2) and proximal femoral fracture detection (M3), including data augmentation and scaling, Efficient-Net and Residual Networks. An ensemble model of 10 models (based on ResNet, VGG, DenseNet, and EfficientNet architectures) was created to detect the presence of a fracture.
Results
With the current images, the accuracy of the developed submodules reached 92-100%, visual explanations of model predictions were generated through gradient-based methods. Time for the automated model-based fracture–labelling was 0.03 seconds/image, compared to an average of 12 seconds/image for human annotation as calculated in our preprocessing stages.
Conclusion
The semi-supervised DL approach showed high accuracy and may mitigate the burden of typically time-consuming and prone to underreporting annotations in constructing a large institutional hip fracture dataset and registry. The developed approach may prove beneficial for future efforts to construct DL-based orthopaedic registries that outperform current diagnosis and procedural codes. Clinicians and researchers can use the developed DL model approach for quality improvement, diagnostic and prognostic research purposes, and building clinical decision support tools.