MTD-GAN: Multi-task Discriminator Based Generative Adversarial Networks for Low-Dose CT Denoising

Abstract

AbstractRadiation dose reduction of computed tomography (CT) is an important research topic due to the potential risk of X-rays. However, low-dose CT (LDCT) images inevitably have a noise that can compromise diagnoses. Recently, although various deep learning algorithms were applied for LDCT denoising, there are still some issues including over-smoothness and visually awkwardness for radiologists. In this paper, we propose a multi-task discriminator based generative adversarial network (MTD-GAN) simultaneously conducting three vision tasks (classification, segmentation, and reconstruction) in a discriminator. To stabilize GAN training, we introduce two novel loss functions termed non-difference suppression (NDS) loss and reconstruction consistency (RC) loss. Furthermore, we take a fast Fourier transform with convolution block (FFT-Conv Block) in the generator to make use of both high- and low-frequency features. Our model has been evaluated by pixel-space and feature-space based metrics in the head and neck LDCT denoising task, and results show outperformance quantitatively and qualitatively than the state-of-the-art denoising methods.KeywordsMulti-task learningGenerative adversarial networkLow-dose CT denoisingHead and neck CT