[论文解读] ILV: Iterative Latent Volumes for Fast and Accurate Sparse-View CT Reconstruction

Seungryong Lee, Woojeong Baek+2 more|arXiv (Cornell University)|Mar 16, 2026

Medical Imaging Techniques and Applications被引用 0

一句话总结

ILV introduces an explicit 3D latent volume refined iteratively by injecting multi-view X-ray features and learned priors, achieving fast and accurate sparse-view CBCT reconstruction and novel-view synthesis. It outperforms prior feed-forward and optimization-based methods while delivering near real-time inference.

ABSTRACT

A long-term goal in CT imaging is to achieve fast and accurate 3D reconstruction from sparse-view projections, thereby reducing radiation exposure, lowering system cost, and enabling timely imaging in clinical workflows. Recent feed-forward approaches have shown strong potential toward this overarching goal, yet their results still suffer from artifacts and loss of fine details. In this work, we introduce Iterative Latent Volumes (ILV), a feed-forward framework that integrates data-driven priors with classical iterative reconstruction principles to overcome key limitations of prior feed-forward models in sparse-view CBCT reconstruction. At its core, ILV constructs an explicit 3D latent volume that is repeatedly updated by conditioning on multi-view X-ray features and the learned anatomical prior, enabling the recovery of fine structural details beyond the reach of prior feed-forward models. In addition, we develop and incorporate several key architectural components, including an X-ray feature volume, group cross-attention, efficient self-attention, and view-wise feature aggregation, that efficiently realize its core latent volume refinement concept. Extensive experiments on a large-scale dataset of approximately 14,000 CT volumes demonstrate that ILV significantly outperforms existing feed-forward and optimization-based methods in both reconstruction quality and speed. These results show that ILV enables fast and accurate sparse-view CBCT reconstruction suitable for clinical use. The project page is available at: https://sngryonglee.github.io/ILV/.

研究动机与目标

• Motivate fast, accurate sparse-view CT reconstruction to reduce radiation exposure and improve clinical workflow.
• Propose a 3D latent volume framework (ILV) that iteratively refines a CT volume using multi-view X-ray features and priors.
• Integrate architectural components (X-ray feature volume, group cross-attention, efficient self-attention, view-wise aggregation) to enable scalable 3D latent refinement.
• Demonstrate via large-scale experiments that ILV surpasses existing feed-forward and optimization-based methods in quality and speed, including novel-view synthesis capabilities.

提出的方法

• Construct an explicit 3D latent volume and iteratively refine it by conditioning on multi-view X-ray features and learned priors.
• Create an X-ray feature volume by back-projecting high- and low-level features from multi-view X-ray images into 3D space.
• Apply group cross-attention between the X-ray feature volume and the latent volume to enable efficient multi-view information injection.
• Use efficient self-attention within the latent volume to capture global context with reduced complexity.
• Incorporate view-wise mean-max features from the X-ray feature volume to reinforce cross-view consistency.
• Decode the refined latent volume into Gaussian primitives and render projections with a differentiable Gaussian renderer, followed by a 3D U-Net refinement for CT volumes.
• Train end-to-end with a composite loss: volumetric MSE to ground-truth, image-space L1/SSIM rendering loss, and a refinement MSE after stabilization.

实验结果

研究问题

• RQ1 Can an explicit 3D latent volume, refined iteratively with learned priors and multi-view X-ray information, achieve superior sparse-view CT reconstruction?
• RQ2 How do architectural choices (group cross-attention, efficient self-attention, view-wise aggregation) affect reconstruction quality and efficiency?
• RQ3 Is the ILV framework capable of high-quality CT reconstruction and reliable novel-view X-ray synthesis under sparse-view conditions?
• RQ4 Do learned priors generalize across datasets and enable near real-time performance compared to optimization-based methods?

主要发现

• ILV achieves state-of-the-art PSNR/SSIM across 6-, 8-, and 10-view settings, with the 10-view PSNR of 33.84 dB and SSIM 0.924 on the main table.
• ILV reconstructs each case in under one second, offering near real-time performance while outperforming both 3D feed-forward and optimization-based baselines.
• Under 24-view settings, ILV surpasses R2-Gaussian by about 2.6 dB in PSNR and is approximately 400x faster than the fastest optimization-based baseline.
• Cross-dataset evaluation shows ILV generalizes better than several baselines when tested on FUMPE and PENGWIN without retraining.
• Ablation studies indicate iterative multi-view injection and the 3D U-Net refinement contribute meaningfully to performance, with full ILV achieving the best results.
• Perceptual fine-tuning with LPIPS reduces perceptual error while maintaining high PSNR/SSIM.