by John R. Fischer
, Senior Reporter | July 12, 2018
Canadian researchers have developed a new AI-based approach for identifying rare pathologies despite a scarcity in the number of X-ray images available for such conditions.
Utilizing machine learning, members of the University of Toronto and St. Michael’s Hospital created computer generated X-rays for the purpose of training AI systems to identify infrequent conditions in authentic X-ray images.
“The X-ray database at St. Mike's was unbalanced. There were more cases of cardiomegaly as compared to pneumothorax. This is consistent with the number of cases in practice,” Professor Shahrokh Valaee, a professor in The Edward S. Rogers Sr. Department of Electrical & Computer Engineering (ECE) at the University of Toronto, told HCB News. “Unfortunately, it is easier for a radiologist to miss those rare cases. By synthesizing X-rays we were able to increase the size of the database and to balance it. We show that the accuracy of detecting those rare cases increases dramatically.”
Carestream Health is a leading provider of quality X-ray systems and detectors that are designed to maximize diagnostic confidence, workflow and patient satisfaction. Follow the link above to see our complete portfolio of digital radiography solutions.
The potential for AI relies on the availability of data in the form of thousands of labeled images. Such amounts though do not exist for certain conditions, preventing speedy and accurate medical diagnostics through machine learning.
The team generated and continually improved upon simulated images utilizing an approach known as deep convolutional generative adversarial network (DCGAN), which is actually composed of two networks. One enables the generation of images while the other tries to discriminate between synthetic and authentic ones. The two are trained to the point where the discriminator is unable to differentiate real from synthesized ones.
Upon obtaining the necessary number of artificial X-rays needed, researchers then combined them with real X-rays to train a deep convolutional neural network to identify normal images and ones depicting a variety of conditions.
Accuracy between the augmented dataset and the original were compared by the Machine Intelligence in Medicine Lab (MIMLab), a group consisting of physicians, scientists and engineering researchers working together in image processing, AI and medicine to solve medical challenges.
Feeding both sets of data through their AI system, the MIMLab found classification accuracy improved by 20 percent for common conditions and 40 percent for rare ones.
Seeing these improvements, the team is confident that artificial X-rays can accurately be used for diagnoses, allowing researchers outside the hospital premises to access them immediately for fast use without violating patient privacy concerns.
GAN, according to Valaee, has been utilized in other areas of healthcare and has just recently been introduced to medical imaging for assessing eye and liver conditions.
“GAN-generated images can be used in retina and liver diseases,” said Valaee. “I expect that many more cases can also benefit from synthesized images.”