A deep learning approach using an ensemble model to autocreate an image-based hip fracture registry

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Informatics

Background:

• Hip fractures are a significant health concern in the elderly, with rising global incidence.
• Existing fracture registries often rely on billing codes, leading to under-reporting and inaccuracies.
• There is a need for more efficient and accurate methods for hip fracture identification and registry creation.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) based approach for automatically creating a hip fracture registry.
• To assess the accuracy and efficiency of DL in detecting hip fractures from radiographic images.
• To provide a tool for quality surveillance and research in hip fracture patient populations.

Main Methods:

• A cascade DL model was designed with three submodules: image view classification, postoperative implant detection, and proximal femoral fracture detection.
• An ensemble model of 10 neural networks (ResNet, VGG, DenseNet, EfficientNet) was trained on 18,834 conventional hip radiographs.
• Data augmentation and scaling techniques were employed to enhance model performance.

Main Results:

• The DL submodules achieved high accuracy, ranging from 92% to 100%.
• Automated fracture labeling by the DL model took only 0.03 seconds per image, significantly faster than manual annotation (12 seconds/image).
• Visual explanations for model predictions were generated using gradient-based methods.

Conclusions:

• The developed semisupervised DL approach accurately labels hip fractures, addressing the limitations of manual annotation.
• This automated method can mitigate the time burden and under-reporting issues associated with large datasets.
• The DL approach offers potential benefits for quality improvement, research, and clinical decision support in hip fracture care.