Paper: Low-rank Angular Prior Guided Multi-diffusion Model for Few-shot Low-dose CT Reconstruction

Authors: Wenhao Zhang, Bin Huang, Shuyue Chen, Xiaoling Xu, Weiwen Wu, Qiegen Liu

The code and the algorithm are for non-comercial use only. Copyright 2024, School of Information Engineering, Nanchang University.

Low-dose computed tomography (LDCT) is essential in clinical settings to minimize radiation exposure; however, reducing the dose often leads to a significant decline in image quality. Additionally, conventional deep learning approaches typically require large datasets, raising con-cerns about privacy, costs, and time constraints. To ad-dress these challenges, a few-shot low-dose CT reconstruc-tion method is proposed, utilizing low-Rank Angular Pri-or (RAP) multi-diffusion model. In the prior learning phase, projection data is transformed into multiple con-secutive views organized by angular segmentation, allow-ing for the extraction of rich prior information through low-rank processing. This structured approach enhances the learning capacity of the multi-diffusion model. Dur-ing the iterative reconstruction phase, a stochastic differ-ential equation solver is employed alongside data con-sistency constraints to iteratively refine the acquired pro-jection data. Furthermore, penalized weighted least-squares and total variation techniques are integrated to improve image quality. Results demonstrate that the re-constructed images closely resemble those obtained from normal-dose CT, validating the RAP model as an effec-tive and practical solution for artifact and noise reduc-tion while preserving image fidelity in low-dose situation.

The training pipeline of RAP

The pipeline for iterative reconstruction stage of RAP

Reconstruction results from 1e4 noise level using different methods

(a) The reference image, (b) FBP, (c) SART-TV, (d) CNN, (e) NCSN++, (f) U-ViT, (g) OSDM, (h) RAP (1).

Train

python main_up.py --config=aapm_sin_ncsnpp_up.py --workdir=exp1_up --mode=train --eval_folder=result1_up
python main_middle.py --config=aapm_sin_ncsnpp_middle.py --workdir=exp1_middle --mode=train --eval_folder=result1_middle
python main_down.py --config=aapm_sin_ncsnpp_down.py --workdir=exp1_down  --mode=train --eval_folder=result1_down

Test

python PCsampling_demo.py