Are Pixel-Wise Metrics Reliable for Sparse-View Computed Tomography Reconstruction?

TL;DR

The paper addresses the mismatch between traditional pixel-wise metrics and clinical usefulness in sparse-view CT reconstruction. It introduces anatomy-aware metrics based on segmentation to quantify structural completeness across four anatomical categories and presents CARE, a diffusion-based enhancement framework that injects segmentation-driven priors into existing reconstructions. CARE is model-agnostic and improves structural fidelity across diverse backbones, achieving substantial gains in large organs, small organs, intestines, and vessels, often unseen by PSNR/SSIM. This approach aligns reconstruction quality with clinical utility, enabling more reliable low-dose CT while offering a scalable, plug-in solution for improving anatomical preservation in practice.

Abstract

Widely adopted evaluation metrics for sparse-view CT reconstruction--such as Structural Similarity Index Measure and Peak Signal-to-Noise Ratio--prioritize pixel-wise fidelity but often fail to capture the completeness of critical anatomical structures, particularly small or thin regions that are easily missed. To address this limitation, we propose a suite of novel anatomy-aware evaluation metrics designed to assess structural completeness across anatomical structures, including large organs, small organs, intestines, and vessels. Building on these metrics, we introduce CARE, a Completeness-Aware Reconstruction Enhancement framework that incorporates structural penalties during training to encourage anatomical preservation of significant structures. CARE is model-agnostic and can be seamlessly integrated into analytical, implicit, and generative methods. When applied to these methods, CARE substantially improves structural completeness in CT reconstructions, achieving up to +32% improvement for large organs, +22% for small organs, +40% for intestines, and +36% for vessels.

Figures (33)

• Figure 1: CARE Improves Structural Completeness in Sparse-View CT Reconstruction.Top: Qualitative comparison of reconstructed CT scans from NeRF and R$^2$-GS, with and without our proposed CARE on clinically important structures. Bottom: Quantitative gains in structural completeness across four reconstruction methods. CARE consistently improves results, confirming its model-agnostic and plug-and-play nature.
• Figure 2: Pixel-wise Metrics Overlook Structural Errors in the Focused Anatomical Structures.A. Example pixel-wise metrics' pitfall. Pixel-wise metrics are insensitive to anatomical preservation failure. B. The four types of organs evaluated by CARE.
• Figure 3: Correlation Between Ground Truth (GT) Based and Segmentator Based Anatomy-Aware Metrics. High correlation scores indicate that the anatomy segmentator is a strong substitute for human expert.
• Figure 4: CARE Framework. Given the frozen anatomy segmentator, autoencoder, and the pretrained latent diffusion model, we then adapt the latent diffusion model to real reconstructed CT scans. CARE can be integrated into any reconstruction method to perform its enhancement capability. The overall training is supervised by three loss terms: the noise-prediction loss $\mathcal{L}_n$, pixel-space reconstruction loss $\mathcal{L}_p$, and anatomy-guidance loss $\mathcal{L}_s$.
• Figure 5: Correlation Between Anatomy-Aware Metrics and Pixel-Wise Metrics. The solid red line is a linear fit; dashed lines denote $\pm1\sigma$ of the residuals. Pearson correlation coefficients ($Corr_{\mathrm{SSIM}}\approx0.16$, $Corr_{\mathrm{PSNR}}\approx0.30$) are low, indicating that better anatomical fidelity does not necessarily yield higher SSIM or PSNR.