Self-supervised inter- and intra-slice correlation learning for low-dose CT image restoration without ground truth

Abstract

Training a convolutional neural network (CNN) to reduce noise in low-dose CT (LDCT) images typically relies on supervised learning, which requires input–target pairs of noisy LDCT and corresponding full-dose CT (FDCT) images. Although previous approaches have shown promising results in LDCT image denoising, it is difficult to acquire clinical datasets of LDCT-FDCT image pairs, which require additional and unnecessary radiation dose delivery to patients. In this paper, we propose a self-supervised learning approach to training a CNN-based denoiser with LDCT images alone. As a means of self-supervision, the proposed approach searches inter-pixel correlation of LDCT images in z-direction as well as in-plane direction. To regularize the CNN-based denoiser, thicker LDCT slices are used as image priors during the self-supervised training process in our approach. For efficient self-supervised learning, we adopt a two-stage training strategy with offline pretraining and online finetuning. The proposed approach is thoroughly evaluated with public and private clinical LDCT datasets. Both image quality measures and clinical assessments indicate that the self-supervised denoising model simultaneously reduces noise level and restores anatomical information in LDCT images from the images alone. The experimental results also show that our online finetuning scheme can improve the denoising performance of supervised learning models as well as self-supervised learning models at test time.