Deep learning can estimate full-dose PET images from scans with significantly lower dosages, according to a new study in the Journal of Digital Imaging. The method may make performing PET scans safer and more affordable.
“There is a desire to reduce the dose of radioactive tracer with which the patients are injected to minimize radiation exposure," wrote Sydney Kaplan, of Washington University School of Medicine in St. Louis, and colleagues. “Unfortunately, lower doses of radioactive tracer result in a significant image quality degradation, so higher doses are generally administered in clinical practice.”
The researchers created a residual convolutional neural network (CNN) which analyzed PET images with one-tenth the dose of full-dose images and subsequently estimated their full-dose image quality. Their method preserved edge and structural details by accounting for them in the loss function during training, according to the study.
In their research, Kaplan and colleagues used whole-body PET image slices of two patients who received 18F-2-deoxyglucose (18FDG). Their CNN was trained on 435 image slices from one patient and 440 slices from the other.
The team determined the high mean structural similarity and peak signal-to-noise ratio values, combined with low root mean square error showed the image quality produced by their model was “more comparable” to true full-dose images, they wrote. The regions of interest stats in the estimated scans were “comparable” to areas in the full-dose images, they added.
Kaplan et al. acknowledged their model must be trained on more patients to gather data representative of the population before it can be applied on a larger scale. But they also suggested their approach could lower PET imaging costs because less radioactive material is needed on a per-scan basis.