Technical Summary: Scaling Medical Imaging Report Generation with Multimodal Reinforcement Learning

Problem Statement

Medical imaging report generation is a critical application for healthcare AI, aiming to produce coherent, clinically meaningful diagnostic reports from images like chest radiographs. While frontier vision-language models have shown promise, they face significant competency gaps in high-value verticals like biomedicine. The paper identifies two primary obstacles in current approaches, particularly those relying on Supervised Fine-Tuning (SFT):

1. Lack of Cross-Institution Generalization: Radiology reporting practices vary widely due to institutional guidelines, departmental conventions, and radiologist writing styles. SFT models tend to inherit the lexical biases and phrasing patterns of their specific training datasets. Consequently, while they may perform well on in-distribution benchmarks, they exhibit substantial degradation when evaluated on unseen institutions or external datasets.
2. Misalignment with Clinical Objectives: SFT objectives primarily optimize for next-word prediction, encouraging surface-level lexical similarity to reference reports rather than alignment with clinically important factual attributes. This leads to overfitting on n-gram-based metrics (e.g., BLEU, ROUGE) that correlate weakly with radiological correctness. The result is a brittleness where models fail to produce faithful, clinically grounded reports that generalize across real-world imaging environments.

Methodology: UniRG Framework

To address these challenges, the authors introduce UniRG (Universal Report Generation), a framework that leverages reinforcement learning (RL) as a unifying mechanism to directly optimize for evaluation metrics designed for end applications.

Model Architecture and Training Data

• Base Model: UniRG-CXR is built upon the open-source Qwen3-VL-8B-Instruct foundation model.
• Data: The model is trained on a massive, diverse corpus comprising over 560,000 studies from 80+ institutions, including MIMIC-CXR, CheXpert Plus, ReXGradient, and IU.
• Inputs: The model conditions on the current image, clinical context (e.g., indication), and, when available, prior study information (both images and reports) to simulate real-world longitudinal workflows.
• Outputs: It generates both the "Findings" and "Impression" sections of a radiology report.

Optimization Strategy: Two-Step RL

Unlike standard SFT+RL pipelines, UniRG employs a novel two-step optimization strategy using GRPO (Group Relative Policy Optimization) to balance multiple clinically grounded objectives:

1. Step 1: RadCliQ-Oriented Optimization: The model is optimized for a weighted composite reward aligned with RadCliQ-v1. This reward aggregates multiple metrics:

   • Rule-based: BLEU
   • Model-based: BERTScore, SembScore, RadGraph-F1
   • Weights: 0.370 (BERTScore), 0.253 (SembScore), 0.377 (RadGraph-F1).
   • Goal: To ensure the model generates outputs that are both lexically and clinically aligned with ground-truth reports.

2. Step 2: Error-Reduction Optimization: Starting from the best checkpoint of Step 1, the model undergoes an additional epoch incorporating CheXprompt, an LLM-based metric that quantifies clinical errors.

   • Reward: 1/(\text{# CheXprompt errors} + 1).
   • Regularization: A KL-regularization term (coefficient 0.03) is applied to prevent excessive deviation from the Step 1 policy, ensuring that gains in factual accuracy do not come at the cost of overall report quality.
   • Goal: To explicitly reduce factual reporting errors and hallucinations.

Key Contributions

The paper outlines three main contributions:

1. UniRG Framework: A multi-objective reinforcement learning framework that yields UniRG-CXR, a single model achieving state-of-the-art (SOTA) performance across multiple datasets and evaluation metrics.
2. Comprehensive Evaluation: A rigorous, clinically grounded evaluation suite spanning report-level metrics, disease-level correctness, cross-dataset/institution/demographic generalization, longitudinal assessment, and human evaluation.
3. Demonstration of Universal Capabilities: Evidence that UniRG-CXR overcomes the specialization and metric-overfitting issues of SFT-based models, producing clinically aligned, generalizable reports across diverse real-world conditions.

Results

Performance on Benchmarks

• ReXrank Leaderboard: UniRG-CXR sets a new overall SOTA on the authoritative ReXrank benchmark, outperforming prior models (e.g., MedVersa, MedGemma, MAIRA-2) by wide margins across four datasets (ReXGradient, MIMIC-CXR, IU X-Ray, CheXpert Plus) and two generation settings (Findings-only and Findings + Impression).
• Metric Universality: The model achieves balanced improvements across diverse metrics, including those it was not explicitly optimized for (e.g., RaTEScore), indicating its gains are not due to overfitting specific metrics.
• Clinical Error Reduction: Using CheXprompt, UniRG-CXR produces substantially more error-free or low-error reports (21.3% with ≤1 error) compared to prior SOTA systems (MedVersa: 16.1%, MedGemma: 3.1%). It also markedly reduces high-error reports (≥4 errors: 14.8% vs. 32.3% for MedVersa).

Generalization and Robustness

• Zero-Shot Generalization: When tested on held-out datasets (IU-Xray and a proprietary dataset) without training on them, UniRG-CXR consistently outperforms prior baselines, demonstrating true generalization across institutions and data distributions.
• Demographic Fairness: The model maintains stable, high performance across gender, age, and race subgroups, with no noticeable performance drops in any demographic category.
• Diagnostic Accuracy: Disease-level F1 scores for condition classification (e.g., Pneumonia, Cardiomegaly) show UniRG-CXR leading across all prevalent diseases.

Longitudinal Capabilities

• Temporal Reasoning: By conditioning on prior images and reports, UniRG-CXR excels in longitudinal report generation. It significantly outperforms the "copy prior report" baseline and models like GPT-5 and MedGemma, which often rely on shortcuts.
• Disease Evolution: The model accurately handles complex temporal changes, including new developments, progression, regression, and no-change scenarios, correctly identifying subtle changes in disease states.

Human Evaluation

In a study involving four board-certified radiologists:

• Preference: UniRG-CXR was the most preferred model, achieving the highest ratings for completeness (3.98) and factual accuracy (4.32).
• Error Rates: It demonstrated the lowest total errors (2.49), with significantly fewer false predictions (0.292) and omissions (0.875) compared to baselines.

Significance and Claims

The paper claims that UniRG-CXR represents a substantial advance in radiology report generation by unifying high performance across datasets, metrics, diagnostic levels, longitudinal setups, and demographic subgroups.

• Universality: Unlike previous models that excel only on isolated benchmarks, UniRG-CXR captures underlying clinical semantics rather than memorizing superficial textual templates. It establishes a robust foundation model capable of maintaining reliability across heterogeneous real-world environments.
• Clinical Alignment: By integrating clinical error signals directly into the RL reward design, the framework bridges the gap between traditional text-generation metrics and radiological practice, prioritizing factual correctness over surface-level similarity.
• Foundation for Future AI: The work highlights the promise of reinforcement learning as a key component for next-generation clinical vision-language models. While currently limited to chest X-rays, the framework is presented as a scalable approach for building reliable, assistive AI systems in radiology.

The authors acknowledge that while the system is a significant step forward, it is not error-free, underscoring the continued necessity of radiologist oversight and validation in real-world deployment.