Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Aligning Findings with Diagnosis: A Self-Consistent Reinforcement Learning Framework for Trustworthy Radiology Reporting

Kun Zhao, Siyuan Dai, Pan Wang, Jifeng Song, Hui Ji, Chenghua Lin, Liang Zhan, Haoteng Tang

TL;DR

The paper addresses the challenge of trustworthy radiology report generation with Multimodal LLMs, where architectural heterogeneity and factual hallucinations hinder clinical deployment. It introduces a two-phase approach: Phase I systematically optimizes the vision encoder–LLM backbone to identify a strong foundation, and Phase II implements a self-consistent reinforcement learning framework with a think–then–summarize generation structure guided by a multi-component reward (consistency, reasoning accuracy, diagnostic accuracy, semantic fidelity, and format compliance) using GRPO. This framework yields state-of-the-art clinical efficacy on MIMIC-CXR with significantly reduced hallucinations by enforcing alignment between the descriptive findings and structured diagnostic outputs. The work demonstrates that explicit reasoning–diagnosis alignment and domain-specific visual representations are crucial for trustworthy radiology reporting, with practical implications for safer clinical translation and future enhancements in connector architectures and consistency-driven training for medical MLLMs.

Abstract

Multimodal Large Language Models (MLLMs) have shown strong potential for radiology report generation, yet their clinical translation is hindered by architectural heterogeneity and the prevalence of factual hallucinations. Standard supervised fine-tuning often fails to strictly align linguistic outputs with visual evidence, while existing reinforcement learning approaches struggle with either prohibitive computational costs or limited exploration. To address these challenges, we propose a comprehensive framework for self-consistent radiology report generation. First, we conduct a systematic evaluation to identify optimal vision encoder and LLM backbone configurations for medical imaging. Building on this foundation, we introduce a novel "Reason-then-Summarize" architecture optimized via Group Relative Policy Optimization (GRPO). This framework restructures generation into two distinct components: a think block for detailed findings and an answer block for structured disease labels. By utilizing a multi-dimensional composite reward function, we explicitly penalize logical discrepancies between the generated narrative and the final diagnosis. Extensive experiments on the MIMIC-CXR benchmark demonstrate that our method achieves state-of-the-art performance in clinical efficacy metrics and significantly reduces hallucinations compared to strong supervised baselines.

Aligning Findings with Diagnosis: A Self-Consistent Reinforcement Learning Framework for Trustworthy Radiology Reporting

TL;DR

The paper addresses the challenge of trustworthy radiology report generation with Multimodal LLMs, where architectural heterogeneity and factual hallucinations hinder clinical deployment. It introduces a two-phase approach: Phase I systematically optimizes the vision encoder–LLM backbone to identify a strong foundation, and Phase II implements a self-consistent reinforcement learning framework with a think–then–summarize generation structure guided by a multi-component reward (consistency, reasoning accuracy, diagnostic accuracy, semantic fidelity, and format compliance) using GRPO. This framework yields state-of-the-art clinical efficacy on MIMIC-CXR with significantly reduced hallucinations by enforcing alignment between the descriptive findings and structured diagnostic outputs. The work demonstrates that explicit reasoning–diagnosis alignment and domain-specific visual representations are crucial for trustworthy radiology reporting, with practical implications for safer clinical translation and future enhancements in connector architectures and consistency-driven training for medical MLLMs.

Abstract

Multimodal Large Language Models (MLLMs) have shown strong potential for radiology report generation, yet their clinical translation is hindered by architectural heterogeneity and the prevalence of factual hallucinations. Standard supervised fine-tuning often fails to strictly align linguistic outputs with visual evidence, while existing reinforcement learning approaches struggle with either prohibitive computational costs or limited exploration. To address these challenges, we propose a comprehensive framework for self-consistent radiology report generation. First, we conduct a systematic evaluation to identify optimal vision encoder and LLM backbone configurations for medical imaging. Building on this foundation, we introduce a novel "Reason-then-Summarize" architecture optimized via Group Relative Policy Optimization (GRPO). This framework restructures generation into two distinct components: a think block for detailed findings and an answer block for structured disease labels. By utilizing a multi-dimensional composite reward function, we explicitly penalize logical discrepancies between the generated narrative and the final diagnosis. Extensive experiments on the MIMIC-CXR benchmark demonstrate that our method achieves state-of-the-art performance in clinical efficacy metrics and significantly reduces hallucinations compared to strong supervised baselines.
Paper Structure (24 sections, 12 equations, 3 figures, 3 tables)

This paper contains 24 sections, 12 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Medical MLLMs
  4. RL for MLLMs
  5. Methodology
  6. Phase I: Systematic Architectural Optimization
  7. Phase II: Self-Consistent Reinforcement Learning Framework
  8. "Reason-then-Summarize" Architecture
  9. Composite Reward Function Design
  10. Loss Function
  11. Experimental Setup
  12. Datasets and Protocols
  13. Evaluation Metrics
  14. Baselines
  15. Implementation Details
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Overview of the proposed framework. Phase 1 performs a systematic exploration to identify the optimal vision--language backbone, while Phase 2 introduces a self-consistent reinforcement learning framework to enforce logical rigor and clinical accuracy.
  • Figure 2: Qualitative comparison of report generation and internal coherence. The top row displays the original chest X-ray, the reports generated by SFT and RL models, and the Ground Truth report. The bottom table compares the structured findings derived from the generated report text ("Report findings") versus the model's explicit structured output ("Answer Label"). The green checkmark indicates that the label is 1, while the red cross indicates that the label does not match the ground truth (0 or None).
  • Figure 3: More cases for qualitative comparison of report generation and internal coherence. The top row displays the original chest X-ray, the reports generated by SFT and RL models, and the Ground Truth report. The bottom table compares the structured findings derived from the generated report text ("Report Label") versus the model's explicit structured output ("Answer Label").