Limb Radiograph Deep Learning Classifier

Purpose: Limb Radiographs present a diagnostic challenge especially when clinical data is nonspecific. therefore, bone lesions in early stages are sometimes difficult to identify and as a result diagnosis and treatment may be delayed.

In recent years, advancements in the Artificial intelligence (AI) field, and specifically Deep Neural Networks (DNN), proved to be a powerful tool that can assist in classifying images. These algorithms are based on vast databases collected and labelled by certified professionals relevant to the studied field.

This study’s objective was to evaluate the use of Deep learning algorithms to develop a decision support tool for primary physicians, which can classify radiographs into normal (i.e., no osseous pathology) and abnormal (i.e., suspected osseous pathology). The main goal of the study is to create a proof of concept for future radiograph decision support classifiers of limb radiographs.

Methods: This retrospective study included patients aged 5-40 with upper and lower extremity radiographs. the dataset was divided into an abnormal group which included radiographs containing bone lesions and anormal control group. Initially 10,000 radiographs matched the original research parameters. After screening by radiologists, a total of 1469 radiographs from 351 patients were included in the study. 973 images (245 patients) were classified as Normal and 496 images (106 patients) were classified as Abnormal. The radiographs were then processed and were used to train several DNNs. The best DNN was chosen based on network sensitivity and accuracy.

Results: The DNN which yielded the best results was found to be MobileNet with sensitivity of 98.6% and accuracy of 91.5%.

Conclusion: This study is an initial proof of concept of the feasibility of a decision support algorithm for primary care physicians in diagnosing limb radiograph abnormality.