AI is revolutionizing how medical images are interpreted, helping medical professionals save time analyzing MR images, CT scans, and X-rays. However, one of the significant challenges deep learning scientists working in the medical community face is the lack of accurate and reliable data to train their neural networks.
To solve this problem, a team of researchers from NVIDIA, the Mayo Clinic, and the Massachusetts General Hospital and Brigham and Women's Health Center for Clinical Data Science developed a deep learning-based model that can generate accurate and reliable synthetic images that can be used for training an AI system. For the first time, researchers are now using generative adversarial networks (GANs) to generate abnormal brain MRIs that can be used to train neural networks.
"Data diversity is critical to success when training deep learning models," the researchers state in a paper published on arXiv. "Medical imaging data sets are often imbalanced as pathologic findings are generally rare, which introduces significant challenges when training deep learning models. We propose a method to generate synthetic abnormal MRI images with brain tumors by training a generative adversarial network."
GANs have been used in medical imaging before to generate a motion model from a single preoperative MRI, upsample a low-resolution image, create a synthetic head CT from a brain MRI, perform medical segmentation, and automatically align different types of MRIs, saving doctors hours. Using an NVIDIA DGX-system, which contains NVIDIA Tesla V100 graphics processing units with the cuDNN-accelerated PyTorch deep learning framework, the researchers trained their GAN on data from two publicly available data sets of brain MRIs. One data set contains thousands of 3D T1-weighted brain MRIs of patients with Alzheimer's disease, the other contains about 200 4D brain MRIs of patients with brain tumors. The team based their image-to-image translation method on the pix2pix model, previously developed by NVIDIA researchers.
According to the researcher team, "This offers an automatable, low-cost source of diverse data that can be used to supplement the training set. For example, we can alter a tumor's size, change its location, or place a tumor in an otherwise healthy brain, to systematically have the image and the corresponding annotation."
Because the images are synthetically generated, there are no patient data or privacy concerns. Medical institutions can easily share data they generate with other institutions, creating millions of different combinations that can be used to accelerate the work. The research team hopes that the model can immediately help deep learning scientists generate new data that can be used to detect abnormalities and, in the end, save lives.
In addition to publishing their paper on arXiv, the researchers have made the code available on GitHub.— Source: NVIDIA