Elbow fractures account for 5-15% of all pediatric fractures and up to 85% of operatively treated fractures. The pediatric elbow anatomy is both unique and variable with age, leading to difficulties in interpreting radiographs by primary care providers and emergency physicians. Radiographic diagnosis is, however, critical to avoiding misdiagnosis and correctly guiding management to avoid growth disturbances and neurovascular compromise.
Machine learning is an application of artificial intelligence whereby algorithms are built to recognize relationships between some data provided (e.g., an image) along with some descriptor (e.g., normal vs fracture). Machine learning algorithms were designed for the automated diagnosis of developmental dysplasia of the hip using AP radiographs as well as hip joint ultrasound images. Lower limb alignment has also received significant attention with algorithms developed for the measurement of hip-knee-ankle angles, femoral anatomic-mechanical angles, and weight bearing lines. Other applications found in the literature include ankle fracture classification, knee osteoarthritis classification and Cobb angle measurement.
With respect to pediatric elbow injuries, machine learning has been employed in several works. England et al. demonstrated the use of deep convolutional neural networks for the detection of traumatic elbow effusions in pediatric patients with good sensitivity and specificity. Convolutional neural networks were also developed for the segmentation of pediatric elbow radiographs to identify and highlight bony structures. Most interesting is the work of Rayan et al. where convolutional neural networks were used in combination with recurrent neural networks for integrating information from multiple radiographic views to emulate radiologist decision making.
In this study, we build on the above to generate a set of machine learning algorithms that can emulate clinical decision making using pediatric elbow radiographs. This will be a natural extension of the currently and rapidly advancing literature and can provide for a tool to be used by non-specialist physicians, especially in under-resourced settings.
To generate a set of machine learning algorithms that can emulate clinical decision making using pediatric elbow radiographs.
This is a retrospective study using previously collected pediatric elbow radiographs, radiology reports and other physician reports. Together, these will be used to develop neural networks for the automated identification of pediatric elbow fractures which can then be employed for the automation practice guidance.
Performance of the neural networks will be compared to that of a trained individual. Comparison will focus on quality of the results in terms of accuracy in particular which we define as reaching the same approach, or one of equivalent validity in the literature, to a particular fracture.