Image: Examples of dataset images presents to networks for classification (Photo courtesy of Max Gordon/ Danderyd Hospital).
A new study suggests that artificial intelligence (AI) deep learning algorithms are on par with humans for diagnosing fractures from orthopedic radiographs.
Researchers at Karolinska Institutet (KI; Solna, Sweden), the Royal Institute of Technology (KTH; Stockholm, Sweden), and Danderyd Hospital (Sweden) extracted 256,000 wrist, hand, and ankle radiographs stored at Danderyd Hospital, classifying them by four variables - fracture, laterality, body part, and exam view. Five deep learning networks were then examined, with the most accurate network benchmarked against a gold standard for fractures.
The deep learning networks were then trained to identify fractures in two thirds of the radiographs under the guidance of the researchers, and then independently analyzed the remaining images, which were completely new to the AI program. Analysis was then compared with that of two senior orthopedic surgeons who reviewed the images at the same resolution as the network. The results showed that all networks exhibited an accuracy of at least 90% when identifying laterality, body part, and exam view.
The final accuracy for fractures was estimated at 83% for the best performing network, which was equivalent to that of senior orthopedic surgeons when they were presented with images at the same resolution as the network. According to the researchers, AI has the potential to do even better with access to greater amounts of data, and they have therefore begun a follow-up study that will include Danderyd Hospital's entire orthopedic archive of over a million high-resolution radiographs. The study was published on July 6, 2017, in Acta Orthopaedica.
“Our study shows that AI networks can make assessments on a par with human specialists, and we hope that we'll be able to achieve even better results with high-res X-ray images,” said senior author Max Gordon, MD, assistant consultant in orthopedics at Danderyd Hospital. “If we can go back to our digital archives, we'll also be able to do extensive research on survival, the development of disease and work capacity - studies that have been impossible to do owing to the amount of data to process.”
Deep learning is part of a broader family of machine learning methods that is based on learning data representations, as opposed to task specific algorithms. It involves artificial neural network (ANN) algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation, with each successive layer using the output from the previous layer as input to form a hierarchical representation.
Royal Institute of Technology