South Korean researchers have used a budding machine-learning technique to generate high-quality structural MR images from amyloid PET scans of dementia patients’ brains. They were then able to quantify cortical amyloid load from these MR-less images, which may open the door to ordering PET scans alone for numerous imaging scenarios in which PET/MR is now a preferred diagnostic pathway.
The Journal of Nuclear Medicine published the research report, by Hongyoon Choi, PhD, and Dong Soo Lee, MD, PhD, of Seoul National University, online Dec. 7.
The machine-learning technique, called generative networking, can be used to yield synthetic yet highly realistic visuals for videogames, industrial planning, consumer product visioning and the like.
Choi and Lee trained their generative network to render realistic structural MR images from florbetapir PET images they obtained from a database of the Alzheimer’s Disease Neuroimaging Initiative.
Using real MR-based quantification as the gold standard, they tested their approach against this and other MR-less quantification methods, including various types of PET.
Their key finding was that the generated MR images created from florbetapir PET data showed visually similar signal patterns to the real MR.
In addition, the mean absolute error of standardized uptake value ratio (SUVR) of cortical composite regions estimated by the generated MR-based method was significantly smaller than other MR-less methods.
Further, the generated MR-based SUVR quantification was the closest to the SUVR values estimated by the real MR-based method.
“Structural MR images were successfully generated from amyloid PET images using deep generative networks,” the authors write. “Generated MR images could be used as template for accurate and precise amyloid quantification. This generative method might be used to generate multimodal images of various organs for further quantitative analyses.”