Medical professionals are hopeful that artificial intelligence will do wonders in the field of medical imaging. But, in the case of studying brain tumors, training data has some in-built problems: it doesn’t have necessary abnormal brain images, due to the fact that they are not common. Nvidia has attempted to solve this problem by developing artificial MRI images with their latest research.

A group of scientists from Mayo Clinic, Nvidia and MGH and BWH Center are going to present their research, work this weekend. Their paper explains how generative adversarial networks (GANs) can be used to develop artificial brain MRI images. GANs are two efficient AI systems that work in opposite – one that develops synthetic data for a particular category, and the second one that recognizes fake data. They both get better by working against each other.

Doctors and researchers can take help from GANs to increase their data sets, particularly for uncommon brain diseases data sets.

“Diversity is critical to success when training neural networks, but medical imaging data is usually imbalanced,” Hoo Chang Shin, senior researcher at Nvidia, briefed. “There are so many more normal cases than abnormal cases, when abnormal cases are what we care about, to try to detect and diagnose.”

Shin and his colleagues will be presenting their paper at MICCAI conference in Spain. Their research work looks for any connection between medical imaging and computer science.

Besides increasing the data sets, Shin and others claim that GANs could help in solving the privacy issues when it comes to getting patient data. Because artificial images are anonymous, and not connected with a specific patient, they safely be used outside the hospital.

The research scientists utilized Nvidia DGX-system with PyTorch deep learning platform powered by cuDNN, for training of GAN using two open source brain MRI’s data sets. One data set had the images of brain suffering from Alzheimer, and the other contained the images of brain tumors.

GAN was separately trained with tumor and brain anatomy labels, implying that researchers can change either the tumor or the brain label to create desired synthetic images – for example, specific size or location of a tumor in the brain.

Nonetheless, Shin clarified that as it is not known exactly how the tumors are developed, the researchers cannot generate tumor images out of nothing – GAN requires at least one real tumor image to begin with.

For improvement in this research, Shin claimed that blind testing should be performed to make sure the quality of artificial images. Moreover, more research should also be done to make sure the identity of patients of original data sets is safe. In the end, the main objective is to simply help doctors acquire more knowledge related to uncommon brain tumors using GAN imaging.