Self-supervised deep learning for joint 3D low-dose PET/CT image denoising

Feixiang Zhao, Dongfen Li, Rui Luo, Mingzhe Liu, Xin Jiang, Junjie Hu

Abstract: Highlights•A two-stage self-supervised framework named MAsk-then-Cycle (MAC) is proposed to enable self-supervised denoising of LDPET/LDCT images. MAC includes a masked autoencoder-based pre-training stage and a self-supervised denoising training stage•A self-supervised denoising training strategy called cycle self-recombination (CSR) is proposed to adapt the characteristics of tomographic imaging•CSR-based unified denoising network for joint denoising of 3D LDPET/LDCT image is proposed.•Signal-guided channel attention module is designed to better disentangle anatomy-dependent noise from entangled noise.•Experimental results demonstrate the generalizability of MAC to joint PET/CT denoising task as well as single modality denoising tasks.