Trending in the latest tech news, identifying rare pathologies in medical images has presented a persistent challenge for researchers, because of the scarcity of images that can be used for training AI systems in a controlled and supervised learning situation.
A new approach has been designed by Professor Shahrokh Valaee and his team. This approach uses machine learning for creating computer-generated X-rays to augmented AI training sets.
Valaee, the professor in the Department of Electrical & Computer Engineering (ECE) at the University of Toronto went on to say, “we are using machine learning itself to do machine learning. We are also developing simulated X-rays that reflect certain rare conditions so that we are able to combine them with real X-rays. This enables us to have a sufficiently large database to train the neural networks for identifying these conditions in other X-rays.”
Valaee is a part of the Machine Intelligence in Medicine Lab (MIMLab). MIMLab is a group of scientists, engineering reserchers and physicians who working towards solving medical challenges by combining their expertise in artificial intelligence, medicine as well as image processing. Valaee says, “AI is capable of providing help in numerous ways in the field of medicine. But to make this possible we need a lot of data. The several labeled images that we need to make these systems work don’t exist for a lot of rare conditions.”
The team uses an AI technique which is known as a convolutional generative adversarial network (DCGAN) for creating artificial X-rays that generate and continuously improve the simulated images.
GANs are a type of algorithm made up of two networks: one that generates the images and the other that tries to discriminate synthetic images from real images. The two networks are trained to the point where the discriminator is unable to differentiate real images from synthesized ones. On the creation of a sufficient amount of artificial X-rays, these are combined with real X-rays for training a deep convolutional neural network. This thus classifies the images into different categories such as normal or identifies numerous conditions.
“We’ve been able to show that the data that is generated artificially with the assistance of a deep convolutional GANs can be used to augment real datasets,” explained Valaee. “This enables in acquring greater quality data for training and thus improves the systems informtion for identifying rare cinsitions.”
When fed through their AI system, the MIMLab compared the accuracy of their augmented dataset to the original dataset and found that the accuracy of the classification improved by over 20 percent for common conditions. The accuracy improved up to about 40 percent for some rare conditions. Since the synthesized X-rays are not from real individuals the dataset is readily available to researchers outside the hospital premises without violating privacy concerns.
“It’s exciting because we’ve been able to overcome a hurdle in applying artificial intelligence to medicine by presenting that these augmented datasets help in improving classification accuracy,” Valaee says. “Deep learning works only if the volume of training data is large enough. Also, this is one way for ensuring we have neural networks that can classify images with high precision.”
For more updates and news, keep reading iTMunch!