Research map / Image generation / Super-resolution

CycleINR: Cycle Implicit Neural Representation for Arbitrary-Scale Volumetric Super-Resolution of Medical Data

CVPR 2024 · Image generation › Super-resolution · approach: Perceptual and real-world SR · 21 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 CycleINR, an implicit neural representation model with local attention latent-code sampling and a cycle-consistent loss for arbitrary-scale volumetric super-resolution of anisotropic CT/MRI, and a new SNLI metric for inter-slice noise inconsistency.

Problem

Anisotropic medical 3D volumes have low inter-slice resolution; existing volumetric super-resolution methods are tied to fixed upsampling factors and generate over-smoothed, inter-slice-inconsistent slices.

Approach

Treat volumetric SR as implicit neural function approximation so one model samples arbitrary scales, add a local attention mechanism over latent-code grid sampling, and enforce a cycle-consistent L1 loss (HR generated from LR, then LR reconstructed from HR) to preserve noise level and reduce over-smoothing.

Key results

• At x5 scale, CycleINR reaches 35.00 PSNR versus 36.99 for SAINT, but lowers LPIPS alex to 0.0464 (vs 0.0996 for SAINT) and SNLI to 0.6050 (vs 1.6996 for SAINT).
• On the MSD liver tumor downstream segmentation at x5, CycleINR gets tumor DSC 0.6881 and NSD 0.5640, the best among compared methods on tumor metrics in Table 3 (it also ranks top for liver NSD at x2).
• Ablation shows the cycle-consistent loss reduces SNLI from 1.1714 to 0.6050 at x5 when combined with LAM, with the LAM alone giving a small LPIPS gain.

Measured effects reported in the paper

Surprising findings

• CycleINR has lower PSNR and SSIM than SAINT, TVSRN, and ArSSR at x5 but better perceptual and slice-consistency scores, and the authors argue PSNR/SSIM favor over-smoothed outputs.

Limitations

• Evaluated only on an in-house chest CT dataset and one public liver dataset; generalization to other modalities is untested.
• PSNR and SSIM are lower than several baselines at larger scales, so gains depend on perceptual and proposed SNLI metrics.
• Tumor segmentation results at x2 and x3 are not uniformly best, and interpolation methods win in some settings.

Experimental setup

Datasets: Private chest CT dataset (204 volumes) Medical Segmentation Decathlon Task03 Liver

Benchmarks: Private chest CT dataset MSD Task03 Liver

Metrics: PSNR SSIM LPIPS SNLI DSC NSD

Baselines: Cubic interpolation Trilinear interpolation TVSRN ArSSR SAINT

Builds on: ArSSR LIIF CycleGAN bilateral filtering

Closest papers by meaning

• Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.85)
• Introducing 3D Representation for Dense Volume-to-Volume Translation via Score Fusion (ICML 2025 · similarity 0.85)
• LiFT: Lifted Inter-slice Feature Trajectories for 3D Image Generation from 2D Generators (NeurIPS 2026 · similarity 0.84)
• Learning General-purpose Biomedical Volume Representations using Randomized Synthesis (ICLR 2025 · similarity 0.84)
• Raptor: Scalable Train-Free Embeddings for 3D Medical Volumes Leveraging Pretrained 2D Foundation Models (ICML 2025 · similarity 0.84)
• Data-Efficient Unsupervised Interpolation Without Any Intermediate Frame for 4D Medical Images (CVPR 2024 · similarity 0.84)