One key to developing an artificial intelligence system or AI is a dataset. To improve accuracy, AI systems require very large samples. But the problem is, some datasets tend to be difficult to find.
Facing this challenge, a research team from Nvidia did not run out of mind to develop an AI system for the purposes of brain imaging. Instead of using the original sample, the research team used artificial or synthetic brain MRI images in three dimensions. These images show the brain with various conditions of cancerous lumps.
"For the first time we can make brain imaging that can be used to train neural networks (in AI systems)," explained senior researcher from Nvidia Hu Chang as reported by Venture Beat.
This AI system was developed using the deep PyTorch learning framework from Facebook and trained using the Nvdia DGX platform. The research team also utilizes the general general adversarial network (GAN) which consists of two parts of the neural network. This neural network consists of a generator that produces samples and a discriminator that functions to distinguish artificial samples and original samples to create convincing MRI images of the abnormal brain.
The research team used two sources of datasets to train GAN, namely the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). Memory and computational limitations force the team to reduce the scan resolution from these two sources from 256x256x108 to 128x128x54. But the research team also continued to use the original image for comparison.
Generators in GAN, take pictures from ADNI and learn to produce synthetic brain scans complete with white matter, gray matter and cerebro spinal fluid. From the BRATS dataset, GAN generators produce full segmentation with tumors.
GAN then quickly notes the scanned images. Because GAN treats brain anatomy and tumors on two different labels, the research team can change tumor size and location or attach it to the results of a healthy brain scan to make artificial brain scans for AI system development purposes.
"This can remove concerns about patient privacy because it produces images anonymously," Chang said.
The research team then trained the machine learing model using original brain scan images and synthetic brain scan images produced by GAN. The results of this training allow the machine learing model trained by the research team to carry out analysis with an accuracy rate of 80 percent, 14 percent more accurate than models that are only trained with original brain scan images only.
"Many radiologists showed excitement after we showed the system. They wanted to use the system to produce more sample images for rare diseases," said Chang.
Facing this challenge, a research team from Nvidia did not run out of mind to develop an AI system for the purposes of brain imaging. Instead of using the original sample, the research team used artificial or synthetic brain MRI images in three dimensions. These images show the brain with various conditions of cancerous lumps.
"For the first time we can make brain imaging that can be used to train neural networks (in AI systems)," explained senior researcher from Nvidia Hu Chang as reported by Venture Beat.
This AI system was developed using the deep PyTorch learning framework from Facebook and trained using the Nvdia DGX platform. The research team also utilizes the general general adversarial network (GAN) which consists of two parts of the neural network. This neural network consists of a generator that produces samples and a discriminator that functions to distinguish artificial samples and original samples to create convincing MRI images of the abnormal brain.
The research team used two sources of datasets to train GAN, namely the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). Memory and computational limitations force the team to reduce the scan resolution from these two sources from 256x256x108 to 128x128x54. But the research team also continued to use the original image for comparison.
Generators in GAN, take pictures from ADNI and learn to produce synthetic brain scans complete with white matter, gray matter and cerebro spinal fluid. From the BRATS dataset, GAN generators produce full segmentation with tumors.
GAN then quickly notes the scanned images. Because GAN treats brain anatomy and tumors on two different labels, the research team can change tumor size and location or attach it to the results of a healthy brain scan to make artificial brain scans for AI system development purposes.
"This can remove concerns about patient privacy because it produces images anonymously," Chang said.
The research team then trained the machine learing model using original brain scan images and synthetic brain scan images produced by GAN. The results of this training allow the machine learing model trained by the research team to carry out analysis with an accuracy rate of 80 percent, 14 percent more accurate than models that are only trained with original brain scan images only.
"Many radiologists showed excitement after we showed the system. They wanted to use the system to produce more sample images for rare diseases," said Chang.
No comments:
Post a Comment