Research map / Label-efficient and robust learning / Self-supervised and contrastive

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

AAAI 2025 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Self-supervised neuroimaging representations · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Spener, an unsupervised INR-based sparse-view CT reconstruction method that embeds the previous iteration's image (via a local image encoder) as a prior within a PnP-HQS iterative framework with BM3D denoising.

Problem

Sparse-view CT reconstruction is highly ill-posed; existing unsupervised INR methods lag behind supervised methods on sparser views and are sensitive to noise, while supervised methods fail out-of-domain.

Approach

Use the imperfect reconstruction from the previous iteration as an image-domain prior: a local image encoder extracts features from the prior image that are fed with hash-encoded coordinates into an MLP, and the INR is optimized with data-consistency plus regularization losses inside a PnP-HQS loop whose denoiser is BM3D.

Key results

• On AAPM 90 views, Spener reaches 37.16 dB PSNR vs 34.61 for supervised RegFormer and 34.31 for SCOPE in the noise-free setting.
• Under low dose (I0=5e5) with 90 views on AAPM, Spener achieves 36.17 dB PSNR, 3.25 dB above RegFormer (32.92).
• On the COVID-19 and CMB-CRC head out-of-domain datasets, Spener outperforms supervised methods (e.g., 40.27 dB vs 38.76 dB for SCOPE on CMB-CRC head 90 views).

Measured effects reported in the paper

Surprising findings

• The supervised RegFormer is best on AAPM 60 views SSIM, and Spener does not win every metric on in-domain data, i.e. it is only comparable to supervised SOTA on in-domain data.

Limitations

• Performance on in-domain data is only comparable to supervised SOTA, not superior across all metrics.
• Relies on a hand-chosen BM3D denoiser and FBP initialization; alternative iterative frameworks and stronger encoders are left for future work.
• Evaluation uses only 10 test slices from one patient per dataset and simulated noise.

Experimental setup

Datasets: AAPM 2016 low-dose CT grand challenge COVID-19 CT dataset CMB-CRC head dataset

Benchmarks: AAPM 2016 low-dose CT grand challenge COVID-19 CT dataset CMB-CRC head dataset

Metrics: PSNR SSIM

Baselines: FBP FBPConvNet RegFormer CoIL SCOPE

Builds on: NeRP SCOPE plug-and-play half-quadratic splitting hash encoding BM3D

Closest papers by meaning

• QN-Mixer: A Quasi-Newton MLP-Mixer Model for Sparse-View CT Reconstruction (CVPR 2024 · similarity 0.86)
• C^2RV: Cross-Regional and Cross-View Learning for Sparse-View CBCT Reconstruction (CVPR 2024 · similarity 0.86)
• SSC-VAE: Structured Sparse Coding Based Variational Autoencoder for Detail Preserved Image Reconstruction (AAAI 2025 · similarity 0.84)
• Learning Image Priors Through Patch-Based Diffusion Models for Solving Inverse Problems (NeurIPS 2024 · similarity 0.84)
• SINR: Sparsity Driven Compressed Implicit Neural Representations (CVPR 2025 · similarity 0.84)
• UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights (NeurIPS 2025 · similarity 0.84)