When are Diffusion Priors Helpful in Sparse Reconstruction? A Study with Sparse-view CT

TL;DR

This paper evaluates when diffusion priors are beneficial for sparse-view CT reconstruction in the context of thoracic fat quantification. It uses an unconditional diffusion model trained on 2D CT slices and a guided-inference framework that enforces projection-consistency, assessing performance across a wide range of projection counts with metrics including PSNR, SSIM, fat content accuracy, and fat dice. The key finding is that diffusion priors excel in extremely sparse regimes but plateau and underperform relative to classical $L_1$ and $L_2$ priors as $n_{proj}$ grows, and they struggle to recover fine low-level detail like vasculature. These results imply substantial radiation-dose reductions are possible with diffusion priors in sparse settings, but also highlight safety considerations and the need for further validation before clinical deployment.

Abstract

Diffusion models demonstrate state-of-the-art performance on image generation, and are gaining traction for sparse medical image reconstruction tasks. However, compared to classical reconstruction algorithms relying on simple analytical priors, diffusion models have the dangerous property of producing realistic looking results \emph{even when incorrect}, particularly with few observations. We investigate the utility of diffusion models as priors for image reconstruction by varying the number of observations and comparing their performance to classical priors (sparse and Tikhonov regularization) using pixel-based, structural, and downstream metrics. We make comparisons on low-dose chest wall computed tomography (CT) for fat mass quantification. First, we find that classical priors are superior to diffusion priors when the number of projections is ``sufficient''. Second, we find that diffusion priors can capture a large amount of detail with very few observations, significantly outperforming classical priors. However, they fall short of capturing all details, even with many observations. Finally, we find that the performance of diffusion priors plateau after extremely few ($\approx$10-15) projections. Ultimately, our work highlights potential issues with diffusion-based sparse reconstruction and underscores the importance of further investigation, particularly in high-stakes clinical settings.

Figures (2)

• Figure 1: Diffusion priors produce convincing results with extremely few projections. We show reconstructions using classical (top) and diffusion (bottom) priors for varying number of projections $n_{proj}$. We show the fat segmentation in purple. Under limited $n_{proj}$, we observe the diffusion method perceptually outperforming the classical priors, but under sufficiently large $n_{proj}$, classical priors produce reconstructions whose fat segmentation's adhere more closely to the ground truth.
• Figure 2: Diffusion priors prevail with extremely few projections but may have wrong content and structure. We plot evaluation metric versus number of projections for classical and diffusion priors. For 206 patients (3 slices/patient), we plot the metric and interquantile range (0.05 and 0.95 quantile, $IQR_{5,95}$) for PSNR, SSIM, fat content accuracy, and fat dice score.