Task-Driven Deep Learning for LDCT Image Denoising

Compared with normal-dose computed tomography (NDCT), low-dose CT (LDCT) images have lower potential radiation risk for patients while suffering from the degradation problem by noise. In the past decades, deep learning-based (DL-based) methods have achieved impressive denoising performances in comparison to traditional methods. However, most existing DL-based methods typically preform training on a specific pairs of LDCT/NDCT images and aim to generalize well on clinical scenarios with LDCT images only. It is a difficult task and challenge, denoising LDCT images with various noise characteristics due to different imaging protocols. We propose a task-driven deep learning framework for LDCT image denoising. Specifically, the variational autoencoder (VAE) is leveraged to learn noise distribution. By utilizing abundant open-source NDCT images as the latent references, we then construct pairs of induced-LDCT (namely pseudo-LDCT)/NDCT images rather than simply using pairs of non-induced-LDCT/NDCT images. Thus, the denoising model can perceive the noise within LDCT images directly. Extensive experiments on LDCT datasets (without NDCT references) show that our proposed framework achieves competitive performances compared with existing DL-based LDCT denoising methods.