CAS-IQA: Assessing Synthetic Angiography Quality

• The paper introduces CAS-IQA, a comprehensive framework that replaces a single CAS with three clinically relevant metrics, achieving notable performance improvements on the CAS-3K dataset.
• CAS-IQA is a framework that utilizes a ViT-based CLIP encoder and the MUST module to align synthetic and reference angiographic images through task-specific sub-metrics.
• The framework outputs Vessel Morphology Consistency (VMC), Vessel Branch Detection (VBD), and Overall Quality (OQ) metrics, each mapped to semantic quality levels for enhanced clinical interpretability.

The term “Correspondence Alignment Score (CAS)” is not formally defined in the vision-language literature for medical image quality assessment. In “CAS-IQA: Teaching Vision-LLMs for Synthetic Angiography Quality Assessment,” the acronym “CAS” refers instead to “Contrast-free Angiography Synthesis” and designates a comprehensive framework for assessing the clinical quality of synthetic angiograms using vision-LLMs (VLMs). The CAS-IQA system does not specify a single metric called the “Correspondence Alignment Score”; rather, it derives clinical validation from three fine-grained, task-specific sub-metrics evaluating structural, anatomical, and global quality properties of generated angiographic images. As such, interpretations of “CAS” as a singular index or formula are incorrect in this context (Wang et al., 23 May 2025).

1. Clarification of “CAS” Terminology

In CAS-IQA, “CAS” denotes the overall framework—Contrast-free Angiography Synthesis Image Quality Assessment—rather than a numerical or explicit image correspondence metric. There is no aggregate “CAS score.” Instead, the system outputs three distinct, clinically relevant metrics, each tailored to an aspect of angiographic image quality. Any reference to “Correspondence Alignment Score (CAS)” as a well-defined, standalone metric is unsupported by this framework.

2. Structure of the CAS-IQA Evaluation Framework

The CAS-IQA framework employs a vision-language architecture designed to incorporate multiple auxiliary imaging modalities as input:

• Inputs: Synthetic (generated) angiography, non-contrast mask, and contrast-enhanced reference images.
• Vision Encoder: A ViT-based CLIP encoder extracts visual tokens. Text prompts are tokenized concurrently.
• Feature Fusion Module: The Multi-path featUre fuSion and rouTing (MUST) module routes features to dedicated branches.
• Metric Branches: Three specialized computation paths are established—one for each task-specific quality metric.
• Prediction Head: Visual tokens, compressed by a Q-Former, along with tokenized textual prompts, are input into an LLM (LLaMA-2) for final quality prediction.

3. Task-Specific Clinical Quality Metrics

Instead of one global “CAS,” CAS-IQA outputs three metrics, each computed on a [0,100] continuous scale and mapped to five semantic quality levels (bad, poor, fair, good, excellent):

Metric	Evaluates	Clinical Interpretation
Vessel Morphology Consistency (VMC)	Structural alignment of major vessels to reference	Ensures vessel shape/continuity
Vessel Branch Detection (VBD)	Correctness/completeness of branch structures vs. reference	Focus on critical side-branches
Overall Quality (OQ)	Aggregate score, penalizing artifacts and integrating VMC/VBD	Radiologist’s confidence metric

VMC is most closely aligned with the notion of “correspondence” as it quantifies agreement between generated and real vessel anatomy. VBD emphasizes fidelity of finer vessel branches. OQ incorporates general imaging quality and clinical usability, with weights assigned to VMC and VBD attentions in its computation.

4. Computation of Metric Branches and Score Mapping

Each metric is computed via a specialized attention-based token fusion path:

• VMC Branch: Aligns global vessel structure through a deformable convolution and cross-attention between generated and reference features.
  • $f_V = \mathrm{DConv}(f_C) - f_M$
  • $$A_\mathrm{VMC} = \mathrm{softmax}\left(\frac{Q_V^\mathrm{VMC} K_G^T}{\sqrt{d}}\right)$$
  • $f_\mathrm{VMC} = A_\mathrm{VMC} V_G$
• VBD Branch: Emphasizes detailed branch structure using a lightweight CNN and similar cross-attention.
• OQ Branch: Computes a weighted fusion of general attention and the VMC/VBD attentions:
  • $$f_\mathrm{OQ} = \left(\alpha \cdot \mathrm{softmax}\left(\frac{Q_G K_G^T}{\sqrt{d}}\right) + \beta \cdot A_\mathrm{VMC} + \gamma \cdot A_\mathrm{VBD}\right) V_G$$
  • Weights satisfy $\alpha + \beta + \gamma = 1$.

Continuous predictions are mapped to discrete levels for instruction tuning, and at inference logits over levels are weighted to recover a scalar score.

5. Experimental Results and Comparative Performance

CAS-IQA was validated on the CAS-3K dataset (3,565 synthetic angiographies with expert annotations). It was benchmarked against classical and modern image quality assessment models, including NIQE, BRISQUE, DBCNN, HyperIQA, ManIQA, Q-Align, and MA-AGIQA. CAS-IQA achieved +0.5–4.1% improvements (PLCC/SRCC) over all baselines for VMC, VBD, and OQ. Ablation of the MUST module resulted in significant performance degradation and slower model convergence (approximately 3× slower), with qualitative outputs demonstrating less reliable vessel correspondence and more artifacts (Wang et al., 23 May 2025).

6. Interpretation and Domain Relevance

The three task-specific metrics represent distinct clinical concerns relevant to interventional radiology:

• VMC: Structural (correspondence) fidelity, critical for vessel-level diagnosis.
• VBD: Completeness of anatomical detail, essential for procedural planning.
• OQ: Integrative measure, encompassing both correspondence and artifact suppression.

The mapping from continuous to semantic levels enhances clinical interpretability. The framework’s architecture is tailored to the data modality and clinical relevance of each sub-metric.

7. Common Misconceptions and Precise Usage

No “Correspondence Alignment Score” metric is introduced or used in the CAS-IQA literature. Interpretations equating "CAS" in "CAS-IQA" with a single numerical score are inaccurate. For evaluations involving alignment or correspondence between synthetic and reference angiograms, researchers should refer to the Vessel Morphology Consistency (VMC) metric, which operationalizes this notion within the CAS-IQA system. No other aggregation or index termed “CAS” is present or implied in the published methodology (Wang et al., 23 May 2025).