Skip to content

DobicVLM: Aligning Chest X-Ray Report Generation with Clinically-Grounded Programmatic Rewards via Group Relative Policy Optimization

🕒 Published (v1): 2026-07-21 11:20 UTC · Source: Arxiv · link

Why this paper was selected

GRPO-trained medical VLM for CXR reports; directly replicable RL recipe for domain-specific VLMs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

DobicVLM fine-tunes MedGemma-4B for chest X-ray report generation using a two-stage pipeline: supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO) with interpretable, rule-based programmatic rewards. The approach enforces clinical constraints—structural formatting, anatomical completeness, semantic faithfulness, and length—without a neural reward model. On blinded expert evaluation over 69 real-world cases, DobicVLM outperforms Gemini 2.5 Flash and matches or exceeds MedGemma-4B on the most clinically critical metrics.

Problem

General-purpose VLMs and SFT-only approaches fail to enforce the rigid structural and anatomical requirements of radiology reporting (e.g., mandatory section headers, anatomical completeness), producing hallucinated findings and formatting failures. RLHF-based alignment is unstable and relies on opaque neural reward models ill-suited for clinical auditability. Standard automated metrics (BLEU, ROUGE) correlate poorly with diagnostic accuracy, making evaluation difficult.

Method

DobicVLM builds on MedGemma-4B (Gemma 3 backbone + SigLIP 400M vision encoder) fine-tuned with LoRA (\(r \ll \min(d,k)\), applied to query/value projections). Training proceeds in two stages:

  1. SFT: Maximizes \(\mathcal{L}_{\text{SFT}} = -\mathbb{E}_{(x,y)}\sum_t \log P_\theta(y^t \mid v, p, y^{<t})\) on ~1,000 de-identified CXR-report pairs for 5 epochs.

  2. GRPO: Generates \(G=6\) candidates per input, computes normalized advantages \(A_i = \frac{r_i - \mu}{\sigma}\), and optimizes a clipped surrogate objective with KL penalty (\(\beta=0.01\), \(\epsilon=0.2\)) against the SFT reference policy.

The reward function \(R_{\text{total}} = \sum \lambda_i R_i\) combines four components: - \(R_{\text{struct}}\): binary regex check for section headers (weight 0.30) - \(R_{\text{check}}\): fraction of anatomical checklist terms present (weight 0.30) - \(R_{\text{overlap}}\): ROUGE-L F1 on Impression sections (weight 0.35) - \(R_{\text{len}}\): soft exponential penalty for length outside \([80, 450]\) tokens (weight 0.05)

Key Contributions

  • Two-stage SFT + GRPO training pipeline adapted for clinical radiology on a modest private dataset (~1,000 pairs) using LoRA for compute efficiency.
  • First application of GRPO to radiology report generation with fully programmatic, rule-based rewards (no neural reward model).
  • Clinically interpretable reward design that directly encodes verifiable standards: regex-based structural adherence, anatomical checklist coverage, ROUGE-L semantic similarity, and length constraints.

Results

Blinded expert evaluation on 69 held-out cases (276 total ratings, 4 board-certified physicians, 5-point Likert scale, reported as acceptance rate):

Criterion Gemini 2.5 Flash MedGemma-4B (Base) DobicVLM (Ours)
Accuracy of Impression 10.7% 26.3% 27.2%
Medical Terminology 82.3% 83.2% 86.5%
Low Hallucination Rate 41.0% 68.8% 65.1%
Report Completeness 70.0% 62.7% 60.2%
Appropriate Referral 36.1% 48.3% 30.9%
  • DobicVLM achieves 2.5Ă— higher impression accuracy than Gemini 2.5 Flash (27.2% vs. 10.7%).
  • Hallucination rate improves substantially over Gemini (65.1% vs. 41.0% acceptance) but regresses slightly vs. MedGemma base (68.8%).
  • Referral and completeness scores are lower than MedGemma base, attributed to reward weighting that deprioritizes these criteria.

Limitations

  • Training set is small (~1,000 cases) and single-institution, limiting generalizability.
  • Evaluation set is 69 cases from one facility; inter-rater variability is unquantified.
  • ROUGE-L as semantic reward is an imperfect proxy for factual accuracy; entity-level rewards (e.g., RadGraph) are not yet incorporated.
  • GRPO's exploration slightly increases hallucinations relative to conservative SFT base (65.1% vs. 68.8%).
  • Referral recommendation quality degrades, as \(R_{\text{overlap}}\) focuses on the Impression section and does not reward the referral component.
  • Absolute acceptance rates remain low (27.2% for impression), positioning the system as a draft-assist tool requiring mandatory radiologist review.

Relevance to Vision-Language Models

This paper demonstrates a practical recipe for domain-specializing a general medical VLM (MedGemma-4B) using GRPO with verifiable, programmatic rewards—a pattern increasingly relevant as the VLM community moves beyond SFT toward RL-based alignment. The programmatic reward design avoids the instability and opacity of neural reward models, making it directly applicable to other high-stakes VLM deployment contexts where auditability is required. The work also highlights a recurring tension in medical VLM evaluation: automated metrics like BLEU/ROUGE are insufficient, and the field needs clinically grounded evaluation protocols (as seen with RadLE and RadCliQ). For researchers tracking VLMs, this is a concrete case study of GRPO-based alignment transferring from mathematical reasoning (its original domain) to multimodal clinical generation.