Crowdsourced X-rays suitable for training AI in orthopedic injuries

The Internet is an acceptable source of images for training algorithms to automatically triage patients with dislocated joints and similar orthopedic emergencies, according to a study published May 28 in Skeletal Radiology [1].

Biomedical engineering grad student Jinchi Wei, MSE, of Johns Hopkins, radiologist Paul Yi, MD, of the University of Maryland and colleagues made the conclusion after mining online radiology repositories for X-rays of 50 dislocated and 50 normal shoulders, elbows, natural hips and replacement hips.

The team used these crowdsourced image datasets to train several convolutional neural networks (CNNs), then tested the algorithms on an external test set of 100 corresponding radiographs (50 dislocated joints, 50 healthy) from three hospitals.

They found the best performing CNNs achieved high areas under the ROC curve for all four joint types.

Further, after creating heatmaps to see which areas the CNNs flagged for clinical decision making, the researchers found the AI competent at focusing on appropriate features in both dislocated and healthy joints.

The authors conclude:

With modest numbers of images, radiographs from the Internet can be used to train clinically generalizable CNNs for joint dislocations. Given the rarity of joint dislocations at many centers, online repositories may be a viable source for CNN training data.”