Exploring the Usage of U-Net-based Deep Learning Networks for Reduction of Brain PET Scan Time in Alzheimer's Disease Patients

Abstract

In medical imaging, positron emission tomography (PET) is an imaging technique that uses radiotracers to tag and investigate biological processes. A high-quality PET scan requires a high dosage of such tracers and/or a long scan time in a PET machine, both of which can be sources of discomfort for the patient. In this work, a potential solution based on deep learning is explored, such that PET scans obtained with shorter scan times can be denoised to minimize image quality loss in brain PET scans of Alzheimer’s disease patients. Using the open ADNI database, 215 brain PET studies of Alzheimer's disease patients using 18F-Florbetapir radiotracer were obtained. Each study contains four sequences of 5-minute scans, and the average of these scans is taken to be the true noiseless image. 203 studies were used to train a U-Net based neural network, using a single 5-minute scan as input and the full 20-minute scan as the ground truth. The U-Net neural network is 18 convolutional layers deep, separated into an encoder and a decoder, where each 2-layer pair in the encoder is concatenated to its corresponding parallel pair in the decoder. The first convolutional layers in the encoder have 64 filters, with the number of filters doubling at each encoding depth up to 1024. The decoder halves the number of filters at each convolutional layer pair, and a final convolution layer collapses the number of filters down to 1, generating the output image. A pipeline was developed to obtain the quantitative metrics for the network’s performance by generating the neural network outputs from single 5-minute scans of the validation studies and obtaining the standardized value uptake ratio values across 56 regions of interest. Combined with a qualitative analysis by a single nuclear medicine physician, the outputs from the neural network with a single 5-minute scan input are comparable both qualitatively and quantitatively to the full 20-minute scans.