[Paper Review] CAP-IQA: Context-Aware Prompt-Guided CT Image Quality Assessment

Kazi Ramisa Rifa, Jie Zhang+1 more|arXiv (Cornell University)|Jan 4, 2026

COVID-19 diagnosis using AI0 citations

TL;DR

CAP-IQA combines medical text priors with instance-level context prompts and causal debiasing to predict CT image quality, achieving state-of-the-art correlation on LDCTIQA 2023 and demonstrating generalization on a large pediatric CT dataset.

ABSTRACT

Prompt-based methods, which encode medical priors through descriptive text, have been only minimally explored for CT Image Quality Assessment (IQA). While such prompts can embed prior knowledge about diagnostic quality, they often introduce bias by reflecting idealized definitions that may not hold under real-world degradations such as noise, motion artifacts, or scanner variability. To address this, we propose the Context-Aware Prompt-guided Image Quality Assessment (CAP-IQA) framework, which integrates text-level priors with instance-level context prompts and applies causal debiasing to separate idealized knowledge from factual, image-specific degradations. Our framework combines a CNN-based visual encoder with a domain-specific text encoder to assess diagnostic visibility, anatomical clarity, and noise perception in abdominal CT images. The model leverages radiology-style prompts and context-aware fusion to align semantic and perceptual representations. On the 2023 LDCTIQA challenge benchmark, CAP-IQA achieves an overall correlation score of 2.8590 (sum of PLCC, SROCC, and KROCC), surpassing the top-ranked leaderboard team (2.7427) by 4.24%. Moreover, our comprehensive ablation experiments confirm that prompt-guided fusion and the simplified encoder-only design jointly enhance feature alignment and interpretability. Furthermore, evaluation on an in-house dataset of 91,514 pediatric CT images demonstrates the true generalizability of CAP-IQA in assessing perceptual fidelity in a different patient population.

Motivation & Objective

• Motivate automatic assessment of CT image quality to reflect radiologists' diagnostic judgments under real-world degradations.
• Propose a CAP-IQA framework that combines textual medical priors with image-specific context prompts.
• Mitigate prompt bias via causal debiasing and dynamic cross-prompt attention.
• Demonstrate superior reliability and generalization on LDCTIQA 2023 and an in-house pediatric CT dataset.

Proposed method

• Use a textual branch to encode medical priors with a frozen PubMedBERT-based prompt embedding.
• Process CT images with a CNN-based encoder to produce a bottleneck feature map and a pooled visual feature f.
• Introduce L image-conditioned context prompts c′ derived from f via an MLP to form instance-adaptive prompts π.
• Apply Dynamic Cross-Prompt Attention (DCPA) to fuse visual and prompt features and produce a fused representation.
• Fuse the DCPA output with encoder features and regresses to a CT IQA score scaled to [0,4].
• Train with mean squared error loss against radiologist-derived ground-truth scores.

Experimental results

Research questions

• RQ1 Can text-derived medical priors guided by prompts be effectively fused with image-specific context prompts to predict CT IQA?
• RQ2 Does dynamic cross-prompt attention improve alignment with radiologist scores over vision-only or text-only baselines?
• RQ3 How well does CAP-IQA generalize across institutions and patient populations (e.g., pediatric CT data)?

Key findings

• CAP-IQA achieves the highest overall LDCTIQA-test score (s = 2.8590) with r = 0.9866, ρ = 0.9775, τ = 0.8949.
• CAP-IQA outperforms the top LDCTIQA leaderboard team (s = 2.7427) by 0.1163, corresponding to a 4.24% improvement in overall correlation.
• Ablation studies show context-guided fusion and DyT normalization provide benefits over alternatives.
• CNN-encoder with PubMedBERT text encoder yields best overall performance among evaluated architectures.
• Evaluation on an in-house pediatric CT dataset (91,514 images) supports generalizability of CAP-IQA to different populations.