Radiology Report Generation for Low-Quality X-Ray Images

1. Problem Statement

Current Vision-Language Models (VLMs) for automated Radiology Report Generation (RRG) achieve strong performance on standard benchmarks but suffer from severe degradation when processing low-quality images.

• The Gap: Existing methods implicitly assume high-quality inputs. In real-world clinical settings (e.g., Emergency Departments, ICUs), X-rays are frequently compromised by noise, motion blur, suboptimal exposure, and patient positioning artifacts.
• The Consequence: Models trained on high-quality data often fail on degraded images, potentially leading to missed diagnoses or incorrect clinical decisions. This is particularly critical because the most severe patient cases (e.g., ICU patients with limited mobility) often yield the noisiest scans.
• Current Limitation: There is a lack of standardized benchmarks and robust training strategies specifically designed to handle these realistic image quality variations.

2. Methodology

The authors propose a comprehensive framework comprising three main components: a data stratification pipeline, a quality assessment agent, and a novel training strategy.

A. Data Stratification & The LRRG Benchmark

To address the lack of quality annotations in existing datasets, the authors constructed the Low-quality Radiology Report Generation (LRRG) benchmark using the MIMIC-CXR dataset.

1. Automated Quality Assessment Agent (AQAA):
   • Built upon Qwen-VL-30B, this agent identifies and stratifies image quality.
   • Inputs: It combines Physics-based metrics (Deviation Index from DICOM metadata for exposure accuracy) and Perceptual IQA metrics (NIQE, MUSIQ, CLIPIQA, TOPIQ) with a CEC/EUR16260-aligned prompt.
   • Output: Classifies images into three regimes: Standard, Mildly Suboptimal, and Severely Suboptimal.
   • Validation: The agent was validated against a "Gold Standard" of 100 manually curated "Retake Pairs" (pre-retake low-quality vs. post-retake high-quality), achieving 99% consistency.
2. Dataset Construction:
   • The dataset was split into three distinct regimes: $D_{std}$ (Standard), $D_{mild}$ (Mildly Suboptimal), and $D_{severe}$ (Severely Suboptimal).
   • Strict Leakage Prevention: Patient-level isolation was enforced to ensure no patient appears in both training and testing sets, guaranteeing unbiased evaluation of generalization.

B. Dual-loop Training Strategy (DTS)

To prevent models from relying on quality-specific spurious cues (e.g., noise patterns) rather than diagnostic features, the authors propose a Bi-level Optimization approach.

• Concept: The goal is to learn a parameter state $\theta$ where the gradient descent direction for high-quality data remains consistent with that of low-quality data.
• Mechanism:
  1. Regime Partitioning: At each iteration, data is randomly partitioned into a Support Set ($S$) and a Query Set ($Q$).
  2. Inner Loop (Virtual Adaptation): The model performs a virtual update on $S$ to obtain temporary parameters $\tilde{\theta} = \theta - \alpha \nabla_\theta L(S; \theta)$.
  3. Outer Loop (Gradient Coherence): The model minimizes the loss on $Q$ using the adapted parameters $\tilde{\theta}$.
• Theoretical Basis: Through Taylor expansion, the outer objective approximates to:
  $$L_{total} \approx L(Q; \theta) - \alpha \langle \nabla_\theta L(Q; \theta), \nabla_\theta L(S; \theta) \rangle$$
  The second term acts as a Gradient Coherence Regularizer. It penalizes updates where the gradients from different quality regimes are orthogonal or conflicting, effectively filtering out noise-related gradients and reinforcing shared diagnostic directions.

3. Key Contributions

1. AQAA & LRRG Benchmark: Development of an automated agent to stratify MIMIC-CXR into quality regimes and the establishment of the first rigorous benchmark (LRRG) specifically for evaluating RRG robustness under realistic image degradation.
2. Dual-loop Training Strategy: A novel bi-level optimization framework using regime rotation and gradient consistency regularization to learn quality-agnostic diagnostic features.
3. Empirical Validation: Extensive experiments demonstrating that the proposed method significantly outperforms state-of-the-art models (including R2GenGPT and Qwen3-VL) on degraded inputs without sacrificing performance on standard high-quality data.

4. Results

Experiments were conducted on four benchmarks: Standard ($T_{std}$), Mildly Suboptimal ($T_{mild}$), Severely Suboptimal ($T_{severe}$), and an Auxiliary real-world set ($T_{aux}$).

• Performance on Degraded Data:
  • On the Mildly Suboptimal set ($T_{mild}$), the proposed R2GenGPT-DTS improved Clinical Precision from 0.353 to 0.411 and F1 Score from 0.291 to 0.321 compared to the baseline.
  • On the Severely Suboptimal set ($T_{severe}$), the model maintained robustness, achieving a RadGraph score of 0.275 (vs. 0.269 for baseline) and significantly better performance on the Auxiliary benchmark ($T_{aux}$), outperforming specialized models like EKAGen.
• Clinical Metrics: The approach showed superior performance in RaTEScore (entity-attribute alignment) and RadGraph (clinical relationship preservation), indicating that the generated reports are not just linguistically fluent but medically accurate even under poor image conditions.
• Generalization: The model learned features invariant to quality shifts, proving that robustness does not require sacrificing performance on ideal inputs.

5. Significance

• Clinical Safety: This work addresses a critical safety gap in medical AI. By ensuring models perform reliably on the "messy" images typical of critical care and emergency settings, it reduces the risk of AI-induced diagnostic errors.
• Paradigm Shift: It challenges the prevailing assumption in medical VLM research that input data is clean and standardized, pushing the field toward more realistic, robust, and equitable AI deployment.
• Methodological Innovation: The use of gradient coherence regularization offers a new direction for training models to be invariant to domain shifts (in this case, image quality) without requiring explicit domain labels during inference.

Limitations: The dual-loop strategy introduces computational overhead during training (though inference is unaffected). The AQAA relies on pre-trained VLMs which may have inherent biases, and current validation is limited to 2D Chest X-rays.