Atb-Net: A Novel Attention-Based Convolutional Neural Network for Predicting Full-Dose From 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 realize high-quality PET images for symptomatic applications, a standard measurement of a radioactive compound must be infused into the understanding, which increments the chance of radiation danger. Be that as it may, diminishing the tracer measurements leads to the expanded commotion, destitute signal-to-noise proportion, and corrupted image quality. To address this issue, profound Deep learning-based strategies have been created to foresee full-dose (FD) from low-dose (LD)images. Be that as it may, profound deep learning strategies commonly utilize the total image as input for training of the network. However, depending on the clinical sign, the center is as it were on specific locales inside the body. When the full image is utilized to train the network, the training of the network can be sub-optimal. In this work, we propose an attention-based convolutional neural network (ATB-Net) to foresee FD from LD PET images. The objective is to prepare a demonstration to memorize the era of FD (standard) PET images from the comparing LD PET images comparing to as it were 5% of the standard dosage by centering exclusively on the target locales within the brain utilizing the Automated Anatomical Labeling (AAL) brain outline. To this conclusion, an altered encoder-decoder U-Net with an extra compartment, which changes over the AAL brain outline into a attention outline, was created. Assessment of the execution of the proposed ATB-Net was performed through comparison with the non-local mean (NLM) filter utilizing different measurements, counting peak Signal-to-Noise Ratio (PSNR), Root Mean Square Error (RMSE), and Structural Similarity Index Measure (SSIM). The PSNR, RMSE, and SSIM for the ATB-Net demonstrate were 38.18±0.78, 0.28±0.05 (SUV), and 0.89±0.02. Thus, they were 10.54%, 20.00%, and 3.49% way better than the NLM filter, individually. In addition, the SUV bias within the Hippocampus and Transient locales was decreased when utilizing the ATB-Net compared to the NLM filter. In expansion to the made strides quantitative precision of ATB-Net compared to NLM, this network is competent of at the same time redressing for partial volume effects taking advantage of the anatomical brain locales. © 2021 IEEE.