An unsupervised sparse-view CT reconstruction framework using combination of iterative deep image prior and ADMM

Sparse views and low-dose scanning reduce radiation exposure computed tomography (CT), but the reconstructed images exhibit severe artifacts and noise due to inadequate view sampling and diminished ray intensity. Recent advances in supervised deep learning (DL) methods have achieved remarkable success in CT reconstruction. However, their reliance on large datasets of paired high-quality and degraded images has constrained their applicability.

In this work, we introduce an unsupervised DL framework called ADMM-DRP, which integrates an untrained neural network with alternating direction method of multipliers (ADMM) iterative reconstruction algorithm.

Specifically, the method employs an untrained neural network as an image generator to optimize the data inconsistency in the Radon domain. To avoid the overfitting phenomenon of traditional deep image prior (DIP)-based methods, we further utilize the ADMM with total variation (TV) regularization continuously update the input of the neural network during the training process.

Experiments on sparse-view and low-dose CT reconstruction tasks demonstrate that the proposed framework outperforms conventional supervised and iterative reconstruction methods in terms of metric and visual quality.

ADMM-DRP reduces the algorithm's reliance on training data, achieves excellent performance in sparse-view and low-dose CT reconstruction, and demonstrates substantial potential for further development in medical imaging.