Researchers Generate Synthetic Brain Scans to Improve AI

Artificial intelligence offers great promise for the medical community. Using medical imaging, medical professionals can train deep learning models and equip AI to detect brain tumors and other abnormalities.

The problem is that researchers must have access to a large data set for comparison, and abnormal brain images, by definition, are uncommon. Additionally, the process of building a large data set is time-consuming and expensive.

Without a large set of patient images, it would be difficult to train AI.

Scientists from Nvidia hope to overcome that challenge with their research on creating synthetic brain MRI’s that can help train neural networks. Working with the Mayo Clinic and the MGH & BWH Center for Clinical Data Science in Boston, the team devised a concept which may provide shareable, reliable data.

Using generative adversarial networks, researchers pit two AI systems against one another: one AI creates a synthetic brain MRI image and the other identifies the fake result. As the systems work against one another, they both improve.

GANs could expand the number of images that doctors and researchers have access to, especially those related to rare brain disease.

The development also addresses challenges like patient privacy, and the challenges inherent in sharing patient images and data.

The GAN images successfully produced abnormal brain tumor MRI’s synthetically, providing an automatable set of data to supplement training.

The technology also allows researchers to manipulate the GAN input to create a more diverse data set. By altering the tumor’s size or placing it in a healthy brain, scientists can modify the data set.

This development allows developers to generate hundreds or thousands of synthetic images.

The team used two publicly available data sets to train its GAN: more than 3,400 images from the Alzheimer’s Disease Neuroimaging Initiative, and 200 4-D images from the Multi-modal Brain Tumor Image Segmentation Benchmark.

The generator was fed images from the ADNI data set, from which it learned to produce synthetic brain scans including white matter, grey matter, and cerebral spinal fluid. Then, using the BRATS data, it generated scans and then annotated them, which would typically take a team of experts’ hours.

The system is not currently able to generate an image from scratch. It must have at least one real image to begin creating synthetic ones.

When the team trained a machine using a combination of real and synthetic scans, it achieved 80 percent accuracy, which was better than training done only with actual data.

Future developments include efforts to protect patient data and ensure the quality of synthetic images, in hopes that healthcare professionals will continue learning about rare brain tumors.