Structure-Adaptive Sparse Diffusion in Voxel Space for 3D Medical Image Enhancement

Abstract: Three-dimensional (3D) medical image enhancement, including denoising and super-resolution, is critical for clinical diagnosis in CT, PET, and MRI. Although diffusion models have shown remarkable success in 2D medical imaging, scaling them to high-resolution 3D volumes remains computationally prohibitive due to lengthy diffusion trajectories over high-dimensional volumetric data. We observe that in conditional enhancement, strong anatomical priors in the degraded input render dense noise schedules largely redundant. Leveraging this insight, we propose a sparse voxel-space diffusion framework that trains and samples on a compact set of uniformly subsampled timesteps. The network predicts clean data directly on the data manifold, supervised in velocity space for stable gradient scaling. A lightweight Structure-aware Trajectory Modulation (STM) module recalibrates time embeddings at each network block based on local anatomical content, enabling structure-adaptive denoising over the shared sparse schedule. Operating directly in voxel space, our framework preserves fine anatomical detail without lossy compression while achieving up to $10\times$ training acceleration. Experiments on four datasets spanning CT, PET, and MRI demonstrate state-of-the-art performance on both denoising and super-resolution tasks. Our code is publicly available at: https://github.com/mirthAI/sparse-3d-diffusion.

• The paper introduces a voxel-space diffusion framework that predicts clean data directly, preserving anatomical details in 3D images.
• It employs sparse timestep scheduling and velocity supervision to significantly accelerate training and improve denoising stability.
• Experimental results on CT, PET, and MRI data show state-of-the-art improvements in PSNR, SSIM, and training efficiency with up to 10× faster convergence.

Structure-Adaptive Sparse Diffusion in Voxel Space for 3D Medical Image Enhancement

Introduction

The paper "Structure-Adaptive Sparse Diffusion in Voxel Space for 3D Medical Image Enhancement" (2604.17773) addresses the challenge of enhancing three-dimensional medical images, particularly in modalities like CT, MRI, and PET. The inherent issues of noise and poor resolution in these images, caused by constraints such as radiation dose, scan duration, and hardware limitations, necessitate effective enhancement techniques to improve diagnostic outcomes. Traditional methods, including slice-by-slice and 2.5D approaches, faced limitations in fully utilizing volumetric data, leading to inconsistencies and compromises in detail preservation. The advent of diffusion models offered a promising alternative, yet scaling them to high-resolution 3D imaging is computationally intensive.

Methodology

The authors propose a framework that operates directly in voxel space, eschewing lossy compressions common in generative models. By leveraging anatomical priors present in degraded inputs, the framework simplifies the diffusion process through sparse sampling, facilitating efficient training without compromising detail recovery. The methodology encompasses several innovations:

• $x_0$-Prediction: Instead of regressing noise or velocity, the model predicts clean data directly, preserving anatomical priors and aligning network outputs with the data manifold.
• Sparse Timestep Scheduling: A compact subset of timesteps is employed, mitigating the redundancy of dense schedules. The framework significantly accelerates training by focusing only on essential steps.
• Velocity Supervision: Stabilization is achieved through velocity space supervision to balance gradient scales and reinforce training stability across sparse schedules.

  Figure 1: Overview of the proposed framework.

Structure-aware Trajectory Modulation

The Structure-aware Trajectory Modulation (STM) module is central to the proposed framework, introducing anatomical awareness within the diffusion process. STM employs multi-structure encoding to extract and integrate structural cues from various spatial dimensions of the input, adapting denoising behavior to local content variability:

• Multi-structure Encoder: Structural representation is derived from tri-planar projections, capturing distinct anatomical patterns across volumetric and 2D orientations.
• Block-wise Trajectory Modulation: Time embedding modulation occurs at each UNet block, permitting detailed anatomical encoding across network depths—facilitating differentiated denoising strengths based on local complexities.

Experimental Results

The experiments span denoising and super-resolution tasks across CT, PET, and MRI datasets, emphasizing the framework's robustness and efficiency. Results consistently demonstrate state-of-the-art performance, with notable gains in PSNR, SSIM, MS-SSIM, and HFEN metrics across all evaluations. The visual reconstructions validate the model's efficacy in recovering sharp anatomical details, highlighting its superiority over existing baselines.

Figure 2: Qualitative denoising results on lung CT and brain PET. Our model preserves sharper, more structurally consistent details than all baselines.

Figure 3: Qualitative $4\times$ super-resolution results on aorta CTA and brain MRI. Our model recovers finer anatomical details with fewer artifacts than all baselines.

Furthermore, the computational advantage is manifest in the convergence curves, illustrating up to $10\times$ faster training without sacrificing enhancement quality.

Figure 4: Training convergence curves for lung CT denoising and aorta CTA $4\times$ super-resolution. Our method reaches the baseline's final performance with up to $10\times$ fewer iterations.

Implications and Future Directions

This research contributes significantly to the domain of 3D medical image processing, providing a pathway to more efficient, anatomically adaptive enhancement approaches. The insights regarding anatomical priors and sparse diffusion trajectories offer avenues for further optimization in medical imaging technologies. The potential extension of these findings to other volumetric generative tasks underlines the versatility and applicability of the framework.

Conclusion

The proposed sparse voxel-space diffusion framework stands out by promoting efficient and high-quality 3D medical image enhancement. Through strategic anatomical encoding and sparse sampling, it achieves remarkable improvements in both computational speed and outcome fidelity. As medical imaging continues to evolve, the methods introduced herein set a foundational precedent for adaptive, efficient processing in clinical and research settings.