Deep residual learning with Anscombe transformation for low-dose digital tomosynthesis

Abstract

Deep learning-based convolutional neural networks (CNNs) have been proposed for enhancing the quality of digital tomosynthesis (DTS) images. However, the direct applications of the conventional CNNs for low-dose DTS imaging are limited to provide acceptable image quality due to the inaccurate recognition of complex texture patterns. In this study, a deep residual learning network combined with the Anscombe transformation was proposed for simplifying the complex texture and restoring the low-dose DTS image quality. The proposed network consisted of convolution layers, max-pooling layers, up-sampling layers, and skip connections. The network training was performed to learn the residual images between the ground-truth and low-dose projections, which were converted using the Anscombe transformation. As a result, the proposed network enhanced the quantitative accuracy and noise characteristic of DTS images by 1.01–1.27 and 1.14–1.71 times, respectively, in comparison to low-dose DTS images and other deep learning networks. The spatial resolution of the DTS image restored using the proposed network was 1.12 times higher than that obtained using a deep image learning network. In conclusion, the proposed network can restore the low-dose DTS image quality and provide an optimal model for low-dose DTS imaging.