Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HiSin: A Sinogram-Aware Framework for Efficient High-Resolution Inpainting

Jiaze E, Srutarshi Banerjee, Tekin Bicer, Guannan Wang, Yanfu Zhang, Bin Ren

TL;DR

This work tackles the memory and compute bottlenecks of high-resolution sinogram inpainting in CT by introducing HiSin, a diffusion-based framework that exploits spectral sparsity and structural heterogeneity. It combines a resolution-guided progressive inference pipeline with two inference-time modules: frequency-aware patch skipping and structure-adaptive denoising, enabling memory-efficient, patch-wise high-resolution inpainting without retraining. Empirical results on TomoBank and LoDoPaB show substantial memory reductions (up to 30.81%) and faster inference (up to 17.58%), while maintaining or surpassing state-of-the-art fidelity at 2048×2048 and 1024×1024 resolutions. The method is compatible with existing optimizations and can be integrated into current CT pipelines to allow practical, high-quality reconstruction under stringent hardware constraints.

Abstract

High-resolution sinogram inpainting is essential for computed tomography reconstruction, as missing high-frequency projections can lead to visible artifacts and diagnostic errors. Diffusion models are well-suited for this task due to their robustness and detail-preserving capabilities, but their application to high-resolution inputs is limited by excessive memory and computational demands. To address this limitation, we propose HiSin, a novel diffusion-based framework for efficient sinogram inpainting that exploits spectral sparsity and structural heterogeneity of projection data. It progressively extracts global structure at low resolution and defers high-resolution inference to small patches, enabling memory-efficient inpainting. Considering the structural features of sinograms, we incorporate frequency-aware patch skipping and structure-adaptive step allocation to reduce redundant computation. Experimental results show that HiSin reduces peak memory usage by up to 30.81% and inference time by up to 17.58% than the state-of-the-art framework, and maintains inpainting accuracy across.

HiSin: A Sinogram-Aware Framework for Efficient High-Resolution Inpainting

TL;DR

This work tackles the memory and compute bottlenecks of high-resolution sinogram inpainting in CT by introducing HiSin, a diffusion-based framework that exploits spectral sparsity and structural heterogeneity. It combines a resolution-guided progressive inference pipeline with two inference-time modules: frequency-aware patch skipping and structure-adaptive denoising, enabling memory-efficient, patch-wise high-resolution inpainting without retraining. Empirical results on TomoBank and LoDoPaB show substantial memory reductions (up to 30.81%) and faster inference (up to 17.58%), while maintaining or surpassing state-of-the-art fidelity at 2048×2048 and 1024×1024 resolutions. The method is compatible with existing optimizations and can be integrated into current CT pipelines to allow practical, high-quality reconstruction under stringent hardware constraints.

Abstract

High-resolution sinogram inpainting is essential for computed tomography reconstruction, as missing high-frequency projections can lead to visible artifacts and diagnostic errors. Diffusion models are well-suited for this task due to their robustness and detail-preserving capabilities, but their application to high-resolution inputs is limited by excessive memory and computational demands. To address this limitation, we propose HiSin, a novel diffusion-based framework for efficient sinogram inpainting that exploits spectral sparsity and structural heterogeneity of projection data. It progressively extracts global structure at low resolution and defers high-resolution inference to small patches, enabling memory-efficient inpainting. Considering the structural features of sinograms, we incorporate frequency-aware patch skipping and structure-adaptive step allocation to reduce redundant computation. Experimental results show that HiSin reduces peak memory usage by up to 30.81% and inference time by up to 17.58% than the state-of-the-art framework, and maintains inpainting accuracy across.

Paper Structure

This paper contains 23 sections, 9 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. HiSin: High-Resolution Sinogram Inpainting
  4. Resolution-Guided Progressive Inference
  5. Frequency-Aware Patch Skipping
  6. Frequency-Aware Patch Pruning
  7. Cosine-Based Patch Blending
  8. Structure-Adaptive Denoising
  9. Complexity Score Computation
  10. Patch-Wise Step Mapping
  11. Evaluation
  12. Experimental Setup
  13. Overall Quantitative and Qualitative Results
  14. Ablation Studies
  15. Compatibility with existing optimizations
  16. ...and 8 more sections

Figures (2)

  • Figure 1: Overview of HiSin. The left illustrates the three-stage resolution-guided pipeline: low resolution followed by mid and high resolutions in a progressive refinement scheme, with the final stage performed patch-wise for detail recovery. At each stage, the upsampled output from the previous resolution is fused with the current input before denoising. The right zooms into the high-resolution stage, where two modules are applied: frequency-aware patch skipping (bypassing low-information patches) and structure-adaptive step allocation (assigning variable denoising steps by patch complexity).
  • Figure 2: Qualitative inpainting results on TomoBank (lines 1 to 2) and LoDoPaB (lines 3 to 4) with random mask (ratio = 0.8) at 1024$\times$1024 resolution. Odd columns and even columns show the sinograms and reconstructed images, respectively.