Deep learning rivals radiologists at spotting wrist fractures on x-rays

Artificial intelligence is growing more accurate each day. And a new study found one such platform is nearing radiologist-level success at detecting a common bone fracture.

That deep learning system, described March 9 in the European Journal of Radiology, can detect and localize wrist fractures about as well as human readers. Notably, the researchers pointed out that their model only required a fraction of the data normally needed to train similar algorithms.

It did stumble when attempting to diagnose fractures on wrist radiographs from outside institutions. But with further tweaking, the team believes it could bolster performance and potentially diagnose many other types of bone breaks.  

“While this study only considered distal radius fractures, it is easily imaginable that deep learning system-assisted detection and localization of other types of fractures is possible and warrants further experiments,” wrote Christian Blüthgen, with University Hospital of Zurich, and colleagues.

For their study, the investigators used 524 wrist x-rays (166 with fractures) to train their artificial intelligence. And to test its overall performance, Blüthgen et al. fed the platform 100 images taken from their Swiss hospital and another 200 from an external test set. A trio of radiologists interpreted the radiographs, highlighting fracture locations and their individual region of interest.

Overall, the deep learning system performed similarly to radiologists. However, the automated approach was not as accurate at assessing fractures on outside images, dropping from an area-under-the-curve score of 0.93 down to 0.81. For radiologists the opposite occurred, and they became more accurate when analyzing exams from a different hospital.    

“This highlights the necessity of training a deep learning-based model on data similar to the intended target data, as well as the need of studies propagating a new deep learning-based method to include a validation experiment to test the model’s capacity to generalize,” the authors noted.

After looking at a heat map to visualize how the AI was reading and interpreting scans, the group found that more than 90% of these “areas of peak activation” aligned with radiologists’ annotations. This means both the AI and radiologists turned to similar points on the image to achieve a diagnosis.

Wrist fractures are among the most common type of fracture and appropriate treatment is “crucial” to avoiding future degenerative diseases, the authors noted. Adding a deep learning-based system to a radiologist's toolbox could prove highly beneficial.

“Following the promising results of this study, future studies will test the applicability of a trained DLS as computer aided diagnostics tool in the setting of the clinical workflow,” the authors concluded.