Unsupervised Self-Prior Embedding Neural Representation for Iterative Sparse-View CT Reconstruction

TL;DR

This work tackles sparse-view CT reconstruction under radiation-dose constraints by introducing Spener, an unsupervised Self-prior Embedding Neural Representation. Spener embeds image-domain priors extracted from imperfect INR reconstructions into an iterative, plug-and-play framework that couples a forward-model-consistent INR with a BM3D denoiser, guided by an image encoder that provides local priors. The method achieves competitive performance with supervised state-of-the-art methods on in-domain data and demonstrates superior robustness to out-of-domain acquisitions and noisy sinograms, outperforming current unsupervised INR approaches. Extensive ablations confirm the benefits of iterative prior updates, the image encoder, and denoiser regularization, highlighting Spener’s practical potential for clinically diverse SVCT scenarios. Code is released at the provided repository.

Abstract

Emerging unsupervised implicit neural representation (INR) methods, such as NeRP, NeAT, and SCOPE, have shown great potential to address sparse-view computed tomography (SVCT) inverse problems. Although these INR-based methods perform well in relatively dense SVCT reconstructions, they struggle to achieve comparable performance to supervised methods in sparser SVCT scenarios. They are prone to being affected by noise, limiting their applicability in real clinical settings. Additionally, current methods have not fully explored the use of image domain priors for solving SVCsT inverse problems. In this work, we demonstrate that imperfect reconstruction results can provide effective image domain priors for INRs to enhance performance. To leverage this, we introduce Self-prior embedding neural representation (Spener), a novel unsupervised method for SVCT reconstruction that integrates iterative reconstruction algorithms. During each iteration, Spener extracts local image prior features from the previous iteration and embeds them to constrain the solution space. Experimental results on multiple CT datasets show that our unsupervised Spener method achieves performance comparable to supervised state-of-the-art (SOTA) methods on in-domain data while outperforming them on out-of-domain datasets. Moreover, Spener significantly improves the performance of INR-based methods in handling SVCT with noisy sinograms. Our code is available at https://github.com/MeijiTian/Spener.

Figures (5)

• Figure 1: Overview of Spener model, including (a) iterative reconstruction using an image embedding neural network $\mathcal{F}_\Phi$, (b) architecture of the image embedding neural network $\mathcal{F}_\Phi$, (c) solving the data fidelity subproblem via the image embedding neural network $\mathcal{F}_\Phi$, and (d) solving regularization subproblem via a denoiser $\mathcal{D}_\sigma$.
• Figure 2: Qualitative results of CT images reconstructed by the compared methods on three datasets. The top two rows show results from the AAPM dataset with 60 views, the middle two rows show results from the COVID-19 dataset with 90 views, and the bottom two rows show results from the CMB-CRC head dataset with 90 views.
• Figure 3: Qualitative results of CT image reconstructed by the compared methods under two dose settings, with both results reconstructed from AAPM dataset with 90 views.
• Figure 4: Qualitative and quantitative results of Spener across different iterations on the AAPM dataset with 90 views.
• Figure 5: Performance curves of Spener with different $\lambda$ settings on AAPM dataset with 90 views reconstruction under low-dose acquisition.