💻 computer science

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

DobicVLM is a vision-language model that leverages Group Relative Policy Optimization with clinically-grounded, rule-based rewards to generate chest X-ray reports that outperform strong baselines like Gemini 2.5 Flash in impression accuracy and medical terminology while ensuring structural and semantic adherence without relying on neural reward models.

Original authors: Thanni Adewuyi, Angelica Obayi, Andem Aniekan, Samuel Okoko, Angel Ezendu, Ephraim Usani, Ademide Animasaun, Philip Chibundu, Christian Maurice, Mary Donald Essien, Oluwaseun Odunsi, Oluwasegun Oguntu

Published 2026-07-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thanni Adewuyi, Angelica Obayi, Andem Aniekan, Samuel Okoko, Angel Ezendu, Ephraim Usani, Ademide Animasaun, Philip Chibundu, Christian Maurice, Mary Donald Essien, Oluwaseun Odunsi, Oluwasegun Oguntuase, Abiodun Adereni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a super-smart robot to be a doctor's assistant. You give it a picture of a patient's chest and ask it to write a report about what it sees. This is the world of Artificial Intelligence in medicine, specifically a field called Vision-Language Models. Think of these models as brilliant students who have read millions of books and seen millions of pictures, but they are still learning how to act like a professional.

The big challenge here is trust. A regular AI might look at a chest X-ray and say, "I see a heart and some lungs," which is true, but a real doctor needs a very specific format: a section for "Findings" (what's wrong) and a section for "Impression" (the final diagnosis). If the AI makes up a disease that isn't there (a "hallucination") or forgets to mention a broken rib, it could be dangerous. The goal isn't just to get the words right; it's to get the medical facts right and follow the strict rules doctors use every day. This paper tackles the question: How do we train a robot to stop guessing and start following the rulebook perfectly?

The Story of DobicVLM: The Robot That Learned to Follow the Rules

Meet DobicVLM, a new AI assistant designed to read Chest X-rays and write medical reports. The researchers who built it realized that simply showing a robot thousands of X-rays and asking it to copy the doctors' notes (a method called Supervised Fine-Tuning) wasn't enough. The robot was learning the style of the reports but was still prone to making things up or missing important details. It was like a student who memorized the font and spacing of an essay but forgot to actually answer the question.

To fix this, the team gave the robot a new kind of training called Group Relative Policy Optimization (GRPO). Imagine you are a teacher grading a class. Instead of just giving a grade to one student's essay, you ask the class to write five different versions of the same essay. Then, you compare them all against a strict checklist of rules. If one student's essay follows the rules perfectly, they get a high score. If another student's essay is creative but breaks the rules (like writing a poem instead of a report), they get a lower score. The robot learns by looking at its own group of attempts and figuring out, "Hey, the one that followed the checklist got the best reward!"

The Magic Checklist: Programmatic Rewards
The secret sauce of DobicVLM is its reward system. Instead of a human teacher grading every single report (which is slow and expensive), the researchers built a digital "checklist" that the robot checks against itself. This checklist has four main rules:

Structure: Did you use the right headers like "Findings" and "Impression"? (Checked by a computer script looking for specific words).
Anatomy: Did you mention the important body parts, like the heart or the lungs? (Checked against a list of required terms).
Truthfulness: Does the report match the actual diagnosis? (Checked by comparing the text to the real doctor's report).
Length: Is the report too short or too long? (Checked to ensure it's not a one-word sentence or a novel).

The robot was trained on 1,000 de-identified chest X-ray reports from a private clinic. After this training, the team tested it on 69 new, unseen cases. They didn't use a computer to grade the results; instead, they asked four real doctors (two radiologists and two medical doctors) to read the AI's reports blindly and rate them on a scale of 1 to 5.

What Did They Find?

The results were a mix of big wins and a few trade-offs.

The Wins:
DobicVLM was the clear winner when it came to getting the diagnosis right. In the doctors' ratings, DobicVLM achieved the highest accuracy for the "Impression" section (the most critical part of the report) at 27.2%. This was better than the base model (MedGemma-4B) at 26.3% and way better than the powerful general AI, Gemini 2.5 Flash, which only scored 10.7%. It also did the best job using correct medical terminology, scoring 86.5%.

The Trade-offs:
However, the robot wasn't perfect. Because the training encouraged the robot to be more "creative" in finding the right answer, it sometimes made up small details that weren't there. The team noticed that DobicVLM had a slightly higher rate of "hallucinations" (making things up) compared to the base model (65.1% acceptance vs. 68.8%). It also scored lower on Report Completeness (60.2%) and Appropriate Referrals (30.9%) compared to the other models.

Why the Trade-off?
The authors suggest this happened because the robot was so focused on getting the main diagnosis right that it sometimes skipped the extra details or the "next steps" (referrals) to stay within the length limits. It's like a student who writes a perfect conclusion but forgets to list all the ingredients in the introduction because they were worried about the word count.

The Bottom Line

DobicVLM shows that you can teach an AI to follow strict medical rules without needing a human to grade every single attempt. By using a transparent, rule-based checklist (the "programmatic rewards"), the team created a model that is much better at giving the correct diagnosis than general AI models.

The paper suggests that this approach is a great step forward, especially for places where resources are limited and you can't afford to hire armies of doctors to train robots. However, the authors are careful to say this isn't a "finished" product ready to replace doctors. The accuracy rates, while the best in the test, still show a gap between the AI and human experts. The robot is currently best suited to act as a drafting tool—a helpful assistant that writes the first version of a report for a human doctor to review and sign off on, rather than a machine that works entirely on its own.

Technical Summary: DobicVLM

Problem Statement

Automated Radiology Report Generation (ARRG) for Chest X-Rays (CXRs) faces a critical gap between experimental models and clinical utility. While general Vision-Language Models (VLMs) can generate text, they frequently fail to meet the high-stakes requirements of radiology, specifically:

Structural Adherence: Failure to follow rigid protocols (e.g., distinct "Findings" and "Impression" sections).
Anatomical Completeness: Omission of critical anatomical structures.
Semantic Faithfulness: Generation of "hallucinations" (fabricated findings) or deviations from standard reporting conventions.
Training Limitations: Standard Supervised Fine-Tuning (SFT) often learns statistical patterns but struggles to enforce strict constraints, leading to formatting failures and factual errors.
Reward Model Opacity: Traditional Reinforcement Learning from Human Feedback (RLHF) relies on neural reward models that are often opaque, unstable, and difficult to audit for clinical safety.

Methodology

The authors propose DobicVLM, a specialized framework that aligns chest X-ray reporting with clinical standards through a two-stage training pipeline using the MedGemma-4B architecture.

1. Model Architecture

Base Model: MedGemma-4B, extending Gemma 3 with a SigLIP 400M vision encoder for X-ray processing.
Efficiency: Utilizes Low-Rank Adaptation (LoRA) applied to query and value projections in attention layers, enabling efficient fine-tuning on limited hardware (Nvidia A100 GPUs).

2. Training Pipeline

The training process consists of two distinct stages:

Stage 1: Supervised Fine-Tuning (SFT): The model is trained on 1,000 de-identified image-report pairs to establish baseline competency. Data augmentation includes rotation, flipping, and contrast adjustments.
Stage 2: Group Relative Policy Optimization (GRPO): The policy is refined using GRPO, a critic-free reinforcement learning algorithm. Unlike RLHF, GRPO does not require a separate neural reward model. Instead, it generates a group of $G$ candidate outputs, assigns rewards based on a programmatic function, and optimizes the policy by contrasting these outputs against the group mean and standard deviation.

3. Clinically-Grounded Programmatic Rewards

A core innovation is the replacement of learned reward models with interpretable, rule-based reward functions. The total reward $R_{total}$ is a weighted sum of four components:

Structural Adherence ( $R_{struct}$ ): Regex verification ensuring the presence of required headers (Indication, Comparison, Findings, Impression).
Anatomical Checklist ( $R_{check}$ ): Calculates the fraction of required anatomical terms (e.g., cardiac silhouette, costophrenic angles) present in the generated text.
Semantic Faithfulness ( $R_{overlap}$ ): Uses ROUGE-L F1 scores on the "Impression" section to ensure alignment with ground truth.
Length Constraint ( $R_{len}$ ): A soft penalty for reports falling outside the [80, 450] token range.

Key Contributions

DobicVLM Framework: A robust pipeline for fine-tuning MedGemma-4B on real-world clinical reports using LoRA and GRPO.
GRPO for Clinical Alignment: One of the first applications of GRPO to radiology, demonstrating that critic-free optimization can effectively enforce structural and anatomical constraints without a separate critic model.
Verifiable Programmatic Rewards: A novel reward function that prioritizes transparent, rule-based verification over learned approximations, ensuring the system is auditable and aligned with Electronic Health Record (EHR) formatting standards.

Experimental Results

The model was evaluated on 69 held-out real-world chest X-ray cases from a private clinical dataset. Evaluation was conducted via blinded expert review by four board-certified Nigerian physicians (two radiologists, two medical doctors) using a 5-point Likert scale across five criteria.

Comparative Performance:
DobicVLM was compared against Gemini 2.5 Flash and the base MedGemma-4B.

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%

Key Findings:

Superior Accuracy: DobicVLM achieved the highest scores in Accuracy of Impression (27.2%) and Medical Terminology (86.5%), outperforming both baselines.
Hallucination Trade-off: While DobicVLM significantly reduced hallucinations compared to Gemini 2.5 Flash (65.1% vs 41.0% acceptance), it showed a slight regression compared to the conservative base MedGemma model (68.8%). The authors attribute this to the exploration inherent in GRPO, which may occasionally introduce non-factual elaborations to satisfy semantic rewards.
Completeness and Referrals: The model showed lower scores in completeness and referrals. The authors note this is a deliberate trade-off; the reward weighting prioritizes accurate, concise reporting over exhaustive anatomical descriptions, which helps reduce the "hallucination for completeness" behavior seen in other models.

Significance and Claims

The paper positions DobicVLM as a step toward transparent alignment in resource-limited settings.

Clinical Utility: The 2.5× improvement in impression accuracy over Gemini 2.5 Flash suggests the model is better suited for clinical deployment, specifically as a draft generation tool requiring mandatory human review or as a teaching tool for residents.
Transparency: By using programmatic rewards rather than neural reward models, the system offers auditability, a critical requirement for medical AI.
Modesty: The authors explicitly acknowledge that the 27.2% acceptance rate for impressions highlights a significant remaining gap to full radiologist standards. They emphasize that the system is not a replacement for radiologists but a supportive tool.
Limitations: The study is constrained by a modest dataset size (~1,000 training, 69 evaluation cases) from a single institution. The authors note that ROUGE-L is an imperfect proxy for diagnostic factuality and call for future work involving entity extraction (e.g., RadGraph) and multi-institutional scaling.