Adapting a low-count acquisition of the bone scintigraphy using deep denoising super-resolution convolutional neural network

Abstract

Purpose: Deep-layer learning processing may improve contrast imaging with greater precision in low-count acquisition. However, no data on noise reduction using super-resolution processing for deep-layer learning have been reported in nuclear medicine imaging.

Objectives: This study was designed to evaluate the adaptability of deep denoising super-resolution convolutional neural networks (DDSRCNN) in nuclear medicine by comparing them with denoising convolutional natural networks (DnCNN), Gaussian processing, and nonlinear diffusion (NLD) processing.

Methods: In this study, 156 patients were included. Data were collected using a matrix size of 256 × 256 with a pixel size of 2.46 mm at 0.898 folds, 15% energy window at the center of the photopeak energy (140 keV), and total count of 1000 kilocounts (kct). Following the training and validation of two learning models, we created 100 images for each 20-test datum. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) between each image and the reference image were calculated.

Results: DDSRCNN showed the highest PSNR values for all total counts. Regarding SSIM, DDSRCNN had significantly higher values than the original and Gaussian. In DnCNN, false accumulation was observed as the total counts increased. Regarding PSNR and SSIM transition, the model using 100-500-kct training data was significantly higher than that using 100-kct training data.

Conclusions: Edge-preserving noise reduction processing was possible, and adaptability to low-count acquisition was demonstrated using DDSRCNN. Using training data with different noise levels, DDSRCNN could learn the noise components with high accuracy and contrast improvement.

Keywords: Bone scintigraphy; Deep denoising super-resolution convolutional neural networks; Gaussian processing; Nonlinear diffusion processing.