Abstract:3043 Objectives: Positron Emission Tomography (PET) is a powerful tool in varies clinical applications including oncology, cardiology and neurology. In PET imaging radioactive tracers are injected to create signals for image generation and it may induce substantial risks to the patients due to radiation exposure. This risk can be mitigated through reducing the dose injected, however the signal to noise ratio of the PET images would also be compromised. In this study we propose a deep learning based technique to improve the SNR of low-counts PET images by combining the information from MRI. Initial results of head images will be demonstrated. Methods: 47 patients were scanned with a whole-body simultaneous PET/MR system (uPMR790, United Imaging Healthcare, Shanghai, China) at Zhongshan hospital Shanghai China. The scan of the head position took about 8 minutes in total. Fully 3D time-of-flight (TOF) List mode PET data was acquired then PET images were reconstructed with TOF-OSEM method and reconstruction parameters were as follow: 20 subsets, 4iterations, 30cm FOV, 128[asterisk]128 matrix, 2.5mm slice thickness, post Metz filter. Varies type MR sequences were used in the head position, including T1-weighted, T2-weighted and EPI images. PET and MRI images were visually checked to make sure there was no misregistration between the two modalities. Then PET rawdata for each patient was randomly down-sampled to half of the original statistics and PET images were reconstructed using the same reconstruction method in the original PET acquisitions. The low-counts PET images were used to simulate the low dose situation and the full-counts PET images were used as the reference dataset. In this study we choose the T1-GRE-FSP-ISO image as the input MRI image along with the low-counts PET images since it has an isotropic image matrix.As shown in figure 1, a 2.5D (5 slices for both PET and MR contrast) dense connected U-net was developed in this study to generate high-counts PET images from noisy PET images and MRI images. Different from conventional U-net, the up- and down- blocks embedded in the U-net were replaced with a dense block (4 densely connected bottle-neck units), and the U depth was set to 4. In the training phase, 1940 samples from 38 patients (two patients with tumors) were feed into model, and 300 epochs with a batch size of 12 were explored to train the model. Other 9 patients with 1 having tumor were used for validation.To evaluate the performance of the model, mean pixel values within ROIs containing 9[asterisk]9[asterisk]9 voxels were calculated across the whole image space, and values from the reference images, low-counts images and the images predicted by the model were compared. Relative errors compared with reference images were calculated and the number of ROIs that with error above certain threshold (2% and 5%) were counted. Results: An example of predicted images are shown in Figure 2. Quantitative results for the test datasets were shown in figure 3. Most predicted images show better quantification accuracy compared to the low-counts images. Conclusions: Initial results show that this method could potentially improve the lesion detectability from low-counts PET images. Further improvements will be done, such as including more training datasets and using a fully 3D model. The benefit of adding MRI images as an input in the model will also be thoroughly studied.

Preserved Edge Convolutional Neural Network for Sensitivity Enhancement of Deuterium Metabolic Imaging (DMI)

Machine Learning-Enabled High-Resolution Dynamic Deuterium MR Spectroscopic Imaging

Review of a New Molecular Imaging Method——Deuterium Metabolic Spectroscopy and Imaging

Improving deuterium metabolic imaging (DMI) signal‐to‐noise ratio by spectroscopic multi‐echo bSSFP: A pancreatic cancer investigation

Accelerating quantitative susceptibility and R2* mapping using incoherent undersampling and deep neural network reconstruction

Deep‐learning‐based super‐resolution for accelerating chemical exchange saturation transfer MRI

Deep Learning-Based Denoising for Magnetic Resonance Spectroscopy Signals.

A Faithful Deep Sensitivity Estimation for Accelerated Magnetic Resonance Imaging

Accelerating Chemical Exchange Saturation Transfer Imaging Using a Model-based Deep Neural Network With Synthetic Training Data

Harmonizing 1.5T/3T Diffusion Weighted MRI through Development of Deep Learning Stabilized Microarchitecture Estimators

Noise‐reduction techniques for 1H‐FID‐MRSI at 14.1 T: Monte Carlo validation and in vivo application

High-field Mr Diffusion-Weighted Image Denoising Using a Joint Denoising Convolutional Neural Network.

Deep learning based multiplexed sensitivity-encoding (DL-MUSE) for high-resolution multi-shot DWI

Attenuation correction and truncation completion for breast PET/MR imaging using deep learning

Noise-reduction techniques for 1H-FID-MRSI at 14.1T: Monte-Carlo validation & in vivo application

Deep Learning-Based Denoising of CEST MR Data: A Feasibility Study on Applying Synthetic Phantoms in Medical Imaging

Using a deep learning prior for accelerating hyperpolarized 13C MRSI on synthetic cancer datasets

Magnetic Resonance Spectroscopy Quantification using Deep Learning

Enabling SENSE accelerated 2D CSI for hyperpolarized carbon-13 imaging

PET Dose Reduction Based on a Deep Convolutional Neural Network in Simultaneous PET/MR A Patient Study