Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scale-Cascaded Diffusion Models for Super-Resolution in Medical Imaging

Darshan Thaker, Mahmoud Mostapha, Radu Miron, Shihan Qiu, Mariappan Nadar

TL;DR

The paper tackles the ill-posed problem of MRI super-resolution by integrating multiscale structure into diffusion priors. It introduces a three-level Laplacian-pyramid framework with scale-specific diffusion priors and a scale-cascaded posterior sampling scheme that solves a sequence of better-conditioned SR tasks from coarse to fine scales, using a modified data-consistency objective. The approach demonstrates improved perceptual quality and reduced inference time on brain, knee, and prostate MRI at 2x and 4x SR, outperforming single-scale DiffPIR and DPS baselines. This work provides a principled, efficient way to unify multiscale reconstruction with diffusion priors for medical image SR.

Abstract

Diffusion models have been increasingly used as strong generative priors for solving inverse problems such as super-resolution in medical imaging. However, these approaches typically utilize a diffusion prior trained at a single scale, ignoring the hierarchical scale structure of image data. In this work, we propose to decompose images into Laplacian pyramid scales and train separate diffusion priors for each frequency band. We then develop an algorithm to perform super-resolution that utilizes these priors to progressively refine reconstructions across different scales. Evaluated on brain, knee, and prostate MRI data, our approach both improves perceptual quality over baselines and reduces inference time through smaller coarse-scale networks. Our framework unifies multiscale reconstruction and diffusion priors for medical image super-resolution.

Scale-Cascaded Diffusion Models for Super-Resolution in Medical Imaging

TL;DR

The paper tackles the ill-posed problem of MRI super-resolution by integrating multiscale structure into diffusion priors. It introduces a three-level Laplacian-pyramid framework with scale-specific diffusion priors and a scale-cascaded posterior sampling scheme that solves a sequence of better-conditioned SR tasks from coarse to fine scales, using a modified data-consistency objective. The approach demonstrates improved perceptual quality and reduced inference time on brain, knee, and prostate MRI at 2x and 4x SR, outperforming single-scale DiffPIR and DPS baselines. This work provides a principled, efficient way to unify multiscale reconstruction with diffusion priors for medical image SR.

Abstract

Diffusion models have been increasingly used as strong generative priors for solving inverse problems such as super-resolution in medical imaging. However, these approaches typically utilize a diffusion prior trained at a single scale, ignoring the hierarchical scale structure of image data. In this work, we propose to decompose images into Laplacian pyramid scales and train separate diffusion priors for each frequency band. We then develop an algorithm to perform super-resolution that utilizes these priors to progressively refine reconstructions across different scales. Evaluated on brain, knee, and prostate MRI data, our approach both improves perceptual quality over baselines and reduces inference time through smaller coarse-scale networks. Our framework unifies multiscale reconstruction and diffusion priors for medical image super-resolution.
Paper Structure (10 sections, 4 equations, 2 figures, 2 tables)

This paper contains 10 sections, 4 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Method
  4. Laplacian Pyramid Construction
  5. Scale-Cascaded Diffusion Prior
  6. Scale-Cascaded Posterior Sampling
  7. Results
  8. Experimental Setup
  9. 2x Super-Resolution
  10. 4x Super-Resolution

Figures (2)

  • Figure 1: Qualitative Results on 2x super-resolution. Across different anatomies, our algorithm is able to fill in fine-grained detail of the degraded image as shown in the difference image prediction $\mathbf{x}^{(1)}$.
  • Figure 2: Qualitative Results on 4x super-resolution. Compared to DiffPIR, our scale-cascaded algorithm better captures finer anatomy details, resulting in higher quality reconstructions. Zoomed in for clarity.