AI Image Enhancement Minimizes Need for Gadolinium

Ingrid Hein

November 28, 2018

A deep-learning algorithm for brain MRI requires only one-tenth the standard dose of gadolinium and does not diminish image quality, researchers report.

"You won't need high doses of a contrast agent in future," said Enhao Gong, PhD, from Stanford University in California.

The effect of the contrast agent is among one of the most controversial topics in radiology. But now, "we can have a quicker, better-quality image without a high dose of gadolinium," he told Medscape Medical News. This is safer and could mean "better patient care and better exams."

Gong explained how he and his colleagues used artificial intelligence to produce an enhanced brain MRI at the Radiological Society of North America 2018 Annual Meeting in Chicago. The technique uses an algorithm to enhance image features, instead of using full-dose contrast agent.

The team assessed 200 patients who each underwent three 3-dimensional T1-weighted inversion recovery fast spoiled gradient-echo (IR-FSPGR) sequences: a precontrast series with no gadolinium, a postcontrast series with a 10% dose of gadolinium, and a standard series with a full dose of gadolinium.

Volumetric visualization of algorithm-enhanced 10% contrast information (Source, RSNA)

The more than 60,000 image pairs generated were processed by a deep convolutional neural network (3D U-Net). This deep-learning technique was trained to create an approximation of the full-dose image using the other two series.

The images were extremely high definition, Gong reported. There are billions pixels in the datasets used for training, and "each pixel can be considered as one sample," he said. "The model really learns from the tremendous number of pixels. It's important to have a large database of images."

Researchers did a fivefold cross-validation to generate results, including the quantitative metrics of signal-to-noise ratio, root mean square error, and structural similarity index to evaluate enhanced contrast improvement, and qualitative metrics to evaluate the deep-learning enhancement. The team also did a noninferiority test to validate dose-reduction viability without losing image quality. Neuroradiologists evaluated the images for contrast enhancement and overall quality.

They found that the dose of gadolinium could be reduced without sacrificing diagnostic quality. In fact, there was little to no difference between the deep-learning-generated image and the full-dose image.

"The human eye cannot see the difference, but our algorithm can detect the subtle differences," said Gong. However, "we are not suggesting you don't need gadolinium at all."

The human eye cannot see the difference, but our algorithm can detect the subtle differences.

The researchers, who worked on the algorithm for 2 years, have formed a company and are waiting to see if the US Food and Drug Administration approves the technique. They have tested it on a number of different scanners to be sure it works. "We tested on the ones we have access to," Gong explained.

Next, the group plans to use a similar technique to assess other parts of the anatomy. "We are working on images of the spine. We will have different protocols and we will train different models, but the general logic is the same," he said.

The MRI technique could be particularly useful for patients with neurologic diseases, such as multiple sclerosis, who currently face more exposure to gadolinium because they are tested frequently, he added. "It could also benefit patients with marginal renal-function issues, where there are questions about whether or not to use gadolinium."

"While the major commercial promise of AI in radiology appears to be on enhanced productivity, these researchers show how machine learning and AI can improve quality, safety, and cost when applied to medical imaging," said Howard Forman, MD, from Yale University in New Haven, Connecticut.

"The opportunities available for similar applications would seem to be enormous and very exciting," Forman said.

Gong and Forman have disclosed no relevant financial relationships.

Radiological Society of North America (RSNA) 2018 Annual Meeting. Presented November 26, 2018.

Comments

3090D553-9492-4563-8681-AD288FA52ACE
Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as:

processing....