CT-DegradBench: A Physics-Informed Benchmark for CT Degradation Detection and Severity Estimation

Technical Summary: CT-DegradBench and SeSpeCT

Problem Statement
Computed Tomography (CT) images are frequently compromised by acquisition artifacts, including noise, blur, streaking, aliasing, and metal artifacts. These degradations arise from diverse sources such as dose reduction, sparse-view sampling, metallic implants, and patient motion, often occurring in combination at varying severity levels. While deep learning has advanced CT restoration tasks (e.g., denoising, metal artifact reduction), the evaluation of these methods remains problematic. Current assessments rely heavily on Image Quality Assessment (IQA) metrics (e.g., PSNR, SSIM) that lack consistent perceptual and clinical validity across diverse degradation types. Furthermore, existing datasets are typically designed for isolated restoration tasks, lacking unified benchmarks that support systematic analysis of degradation detection, severity estimation, and metric reliability under controlled single- and mixed-artifact conditions.

Methodology
The authors address these gaps through two primary contributions: a new dataset/benchmark and a novel estimation framework.

1. CT-DegradBench (Dataset and Benchmark):

   • Generation: The dataset is constructed using physics-informed degradation models applied to reference CT images from the Mayo 2016 Low-Dose CT Grand Challenge. Degradations are simulated in the projection (sinogram) domain or by modifying attenuation maps before reconstruction via Filtered Backprojection (FBP).
   • Degradation Types: It covers five common artifact families: Noise (mixed Poisson–Gaussian), Loss of Sharpness (blur), Streaking, Aliasing (sparse-view), and Metal artifacts.
   • Severity and Mixtures: Each type is calibrated across four severity levels (L0–L3). Crucially, the benchmark includes "mixed-artifact" settings (e.g., Blur + Noise, Metal + Noise) to approximate clinically plausible compound degradations.
   • Metadata: Each sample includes structured metadata (degradation type, severity, mixture order) and radiologist-informed natural language prompts to facilitate Vision-Language Model (VLM) benchmarking.

2. SeSpeCT (Semantic–Spectral CT Degradation Estimation):

   • Architecture: SeSpeCT is a multi-task framework designed for joint artifact type classification and severity regression. It operates without task-specific fine-tuning of the underlying foundation models.
   • Semantic Quality Branch: Utilizing a frozen medical Vision-Language Model (BioMedCLIP), the method constructs a "semantic quality axis" in the embedding space. This is derived from the normalized difference between embeddings of high-quality and low-quality radiology-informed text prompts. Input CT slices are projected onto this axis to generate global and local semantic quality scores.
   • Spectral Branch: Recognizing that certain artifacts (e.g., aliasing, streaks) manifest distinctly in the frequency domain, a spectral branch extracts features from the Fourier magnitude spectrum using a lightweight CNN and computes a high-frequency energy ratio.
   • Fusion and Prediction: Semantic and spectral descriptors are concatenated and projected via an MLP. The model is trained with a multi-task objective combining cross-entropy classification, SmoothL1 regression, pairwise ranking loss (to preserve ordinal severity structure), and supervised contrastive loss.

Key Results

• Benchmarking IQA and VLMs: The authors evaluated classical IQA metrics (PSNR, SSIM, VIF, LPIPS, DISTS) and frozen VLMs (OpenCLIP, MedCLIP, BioMedCLIP, Merlin) on CT-DegradBench. Results indicate that standard metrics are not equally sensitive to all degradations; for instance, PSNR is responsive to noise but less so to metal artifacts, while structural metrics better capture blur. Similarly, while some VLMs show monotonic correlation with severity for certain artifacts (blur, streaks), they struggle with metal artifacts and mixed degradations.
• SeSpeCT Performance: SeSpeCT consistently outperformed frozen VLM baselines and classical metrics in both single and mixed degradation settings.
  • Severity Estimation: SeSpeCT achieved a mean Spearman correlation of 0.780 for single degradations and 0.710 for mixtures, significantly outperforming the best baseline (BioMedCLIP at 0.665). Notable improvements were seen for metal artifacts, where correlation improved from <0.12 to 0.435.
  • Classification: The model achieved high accuracy (0.783) and F1 scores (0.776) for degradation classification.
  • Ablation: Removing the semantic branch caused a catastrophic drop in performance (Accuracy to 0.151), highlighting the necessity of the prompt-derived semantic prior. Removing the spectral branch also degraded performance, confirming the complementary nature of frequency-domain cues.

Significance and Claims
The paper claims to provide the first unified benchmark for analyzing how different degradations and severity levels affect CT image quality, addressing the lack of standardized evaluation for degradation detection and severity estimation. By introducing CT-DegradBench, the authors enable systematic evaluation of degradation sensitivity and metric reliability under controlled conditions, including complex mixed-artifact scenarios often found in clinical practice.

The proposed SeSpeCT framework demonstrates that combining training-free semantic priors from medical VLMs with frequency-domain features yields a robust representation for CT degradation analysis. The authors assert that this approach offers a more reliable and monotonic representation of degradation severity than existing metrics or frozen VLM features alone, particularly for localized and compound artifacts. The work aims to support more reliable quality assessment and evaluation protocols for future CT restoration and analysis pipelines.