Extraction of Diagnostic Positron Emission Tomography (PET) Images From 10 Seconds Bed-position Acquisition, Using Deep Neural Networks

Positron emission tomography (PET)/ computerized tomography (CT), with the use of several tracers, among which fluoro deoxyglucose (FDG) is the most prevalent, has become a principal imaging modality in oncology. The PET and CT components reflect metabolic and anatomic information, respectively. PET images are based on detecting two annihilation 511 KeV photons that are produced by positron emitting isotopes. The longer the acquisition time, the more photons are detected and processed, resulting in better image quality. However, long scan times (typically 20-40 minutes per scan) are less convenient to patients, and may result in patient motion and misalignment. Over the years, several methods, such as 3D and time of flight acquisitions, have been developed to compensate for the degradation in image quality as a result of shortening of the scanning time. Recently, several studies have used machine learning to produce diagnostic images from low quality images. Xiang et al compared PET images of the brain that were acquired in 3 minutes (i.e., low-quality PET (LPET)) with standard PET images (i.e., SPET) that were acquired in 12 minutes. They have combined LPET and T1 weighted images using deep neural networks (DNN) to produce diagnostic PET images equivalent to SPET images.

The goal of our study is to produce diagnostic PET images with 10 seconds acquisition time per bed position using DNN algorithms developed at the CILAB laboratory in the imaging department of Sheba.

The algorithms were previously successfully validated for the denoising of ultra-low dose chest CT scans, making them suitable for lung cancer screening. The algorithms are based on the locally-consistent non-local means (LC-NLM) algorithm.

The LC-NLM algorithm uses fast approximate nearest neighbors (ANN) to find the most similar high-SNR patch, in a purposely built database, for each noisy patch in the input image (Green et al.) ] We propose to use the recently introduced non-local neural networks (Wang et al.) in order to stack the LC-NLM into a fully trainable, locally-consistent nonlocal block (LC-NLB). The original non-local networks combines the ideas of the classical non-local means (NLM) algorithm (Buades et al.) into a neural network block, which computes the output at a specific position as a weighted sum of the features at all positions.