A Novel Attention-Based Convolutional Neural Network for Joint Denoising and Partial Volume Correction of Low-Dose Pet Images Publisher



Azimi MS¹ ; Kamaliasl A¹ ; Ay MR² ; Arabi H³ ; Zaidi H³ Show Affiliations
Source: 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference Record# NSS/MIC 2021 and 28th International Symposium on Room-Temperature Semiconductor Detectors# RTSD 2022 Published:2021

Abstract

To accomplish an excellent PET image for symptomatic purposes, a standard dose of radioactive tracer should be infused into the patient's body, which builds the danger of radiation harm. Nonetheless, lessening the tracer dose would prompt low quality of the PET image and noise induced quantitative bias. One more worry for precise quantitative PET imaging is partial volume effect (PVE) which are a result of the innately restricted spatial resolution of PET and present huge biases particularly for structures with a similar size request of the framework point spread function (PSF). In this work, we intend to utilize deep learning-based strategies to propose a half and half methodology for foreseeing full-dose and PVE-corrected (FD+PVC) images from the low-dose (LD) partners. deep learning strategies by and large utilize the entire image as contribution for network learning, while the most well-known PVC techniques like GTM (geometric transfer matrix) targeting separating PVE-corrected movement fixation esteems for client characterized volumes of interest (VOIs), so the attention is just on the particular spaces of the image, and when the entire image is utilized, the preparation of the network might be problematic. In this work, we propose an attention based convolutional neural network (ATB-Net) to foresee full-dose and PVE-corrected (FD+PVC) from low-dose (LD) PET images by means of concentrating of the network on the Automated Anatomical Labeling (AAL) brain areas. To assess the presentation of the proposed ATB-Net, the examination was performed with U-Net utilizing PSNR (Peak Signal-to-Noise Ratio) and RMSE (Root Mean Square error) measurements. The quantitative investigation exhibited the prevalent presentation of the proposed ATB-Net with PSNR and RMSE of 20.91% and 12.39% better than those of U-Net, separately. The area shrewd examination showed that the distinction between the relative biases acquired from the U-Net and ATB-Net were all measurably huge (p<0.05). On account of absolute relative bias, the greater part of the locales has critical contrasts (p<0.05). Along these lines, the proposed ATB-Net would have the option to all the while performed correction for the partial volume effects and smother the noise owing from admittance to the anatomical areas of the brain. © 2021 IEEE.