Research map / Reinforcement learning / RL from feedback
Radiology Report Generation via Multi-objective Preference Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MPO, a multi-objective RL method for radiology report generation that conditions a Transformer on a preference vector (via a PVF fusion module) and optimizes preference-weighted NLG and clinical-efficacy rewards with REINFORCE, so one model serves many radiologist preferences.
Problem
Radiology report generation models are trained with supervised regression or single fixed reward mixtures, so they cannot adapt to heterogeneous radiologist preferences (e.g., fluency vs. clinical accuracy) without retraining.
Approach
Encode a preference simplex vector as a condition fused with visual features by an attention-plus-residual PVF module, and train the generator with REINFORCE on linearly weighted multi-dimensional rewards sampled across the preference space, so a single model can be steered at inference without fine-tuning.
Key results
- On MIMIC-CXR, MPO reaches CE F1 0.353 vs. 0.276 for R2Gen base and precision 0.436 (best), while BLEU-1 is 0.416 vs 0.353 for R2Gen.
- On IU-Xray, MPO achieves BLEU-1 0.548 and ROUGE-L 0.415, best among compared methods on all NLG metrics except METEOR.
- Varying preference weights on IU-Xray moves BLEU-1 from 0.548 to 0.531 while ROUGE-L rises from 0.413 to 0.415; on MIMIC-CXR, F1 rises from 0.363 to 0.436 as weight shifts toward F1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MPO multi-objective preference optimization to R2Gen | R2Gen · IU-Xray | BLEU-1: 0.470 -> 0.548 | Helped |
| add MPO multi-objective preference optimization to R2Gen | R2Gen · MIMIC-CXR | F1: 0.276 -> 0.353 | Helped |
| increase BLEU-1 reward weight from 0 to 1 instead of prioritizing ROUGE-L | MPO · IU-Xray | BLEU-1: 0.531 -> 0.548 | Helped |
| increase BLEU-1 reward weight from 0 to 1 instead of prioritizing ROUGE-L | MPO · IU-Xray | ROUGE-L: 0.415 -> 0.413 | Hurt |
| increase F1 reward weight from 0 to 1 instead of prioritizing BLEU-1 | MPO · MIMIC-CXR | F1: 0.363 -> 0.436 | Helped |
| increase F1 reward weight from 0 to 1 instead of prioritizing BLEU-1 | MPO · MIMIC-CXR | BLEU-1: 0.416 -> 0.392 | Hurt |
| add MOO to the Base model (ablation) | R2Gen · IU-Xray | BLEU-1: 0.470 -> 0.540 | Helped |
| add MOO to the Base model (ablation) | R2Gen · MIMIC-CXR | BLEU-1: 0.353 -> 0.381 | Helped |
| add PVF to Base+MOO (ablation) | R2Gen with MOO · IU-Xray | BLEU-1: 0.540 -> 0.548 | Helped |
| add PVF to Base+MOO (ablation) | R2Gen with MOO · MIMIC-CXR | BLEU-1: 0.381 -> 0.416 | Helped |
| replace PVF attention fusion with concatenation (ablation) | MPO · IU-Xray | BLEU-1: 0.548 -> 0.544 | Hurt |
| replace PVF attention fusion with addition (ablation) | MPO · IU-Xray | BLEU-1: 0.548 -> 0.541 | Hurt |
Surprising findings
- Adding the PVF preference fusion to MOO improves over MOO alone on all metrics, while simpler fusions (concat, add, mul) perform worse than no fusion gain.
- MAN reports higher CE recall and F1 (0.398, 0.389) than MPO on MIMIC-CXR, so MPO is not best on all clinical metrics.
Limitations
- Only two-dimensional preference spaces are reported in the main text; three-dimensional results are deferred to supplementary material.
- Rewards are limited to existing NLG and CE metrics; the authors note designing new evaluation metrics as rewards is future work.
- Radiologist preferences are modeled as metric weight vectors, not collected from actual radiologists; no human evaluation is reported.
Experimental setup
Datasets: IU-Xray MIMIC-CXR
Benchmarks: IU-Xray MIMIC-CXR
Metrics: BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE-L CE precision CE recall CE F1
Baselines: R2Gen CMN CA CMN+RL XPRONET KiUT METrans MMTN MAN COMG+RL
Builds on: R2Gen CMN+RL REINFORCE ResNet
Closest papers by meaning
- ReFINE: A Reward-Based Framework for Interpretable and Nuanced Evaluation of Radiology Report Generation (AAAI 2026 · similarity 0.88)
- Bootstrapping Large Language Models for Radiology Report Generation (AAAI 2024 · similarity 0.87)
- LLM-RG4: Flexible and Factual Radiology Report Generation Across Diverse Input Contexts (AAAI 2025 · similarity 0.86)
- Reporting Practice Matters: The Impact of Reference Choice on Chest X-ray Report Evaluation (NeurIPS 2026 · similarity 0.85)
- Instance-level Expert Knowledge and Aggregate Discriminative Attention for Radiology Report Generation (CVPR 2024 · similarity 0.84)
- PromptMRG: Diagnosis-Driven Prompts for Medical Report Generation (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.