DeepMedix-R1: Chest X-ray Reasoning Model

• DeepMedix-R1 is a chest X-ray foundation model that integrates image grounding with sequential online reinforcement learning for transparent clinical interpretation.
• It employs a three-stage training pipeline—supervised tuning, synthetic reasoning initialization, and online reinforcement learning—to refine its performance.
• Evaluation shows significant gains in report generation and visual question answering, with improvements up to 57.75% over existing models.

DeepMedix-R1 is a chest X-ray foundation model for image-grounded medical reasoning that is designed to generate both clinical answers and explicit reasoning steps tied to local image regions. Introduced as “A Foundation Model for Chest X-ray Interpretation with Grounded Reasoning via Online Reinforcement Learning,” it specializes Qwen2.5-VL-7B for chest radiography through a sequential pipeline consisting of supervised instruction tuning, synthetic cold-start reasoning initialization, and online reinforcement learning. Its stated aim is to move chest X-ray interpretation away from black-box generation toward transparent, locally grounded, and clinically actionable output across report generation and visual question answering tasks (Lin et al., 4 Sep 2025).

1. Definition, scope, and problem setting

DeepMedix-R1 is presented as a holistic medical foundation model for chest X-ray interpretation rather than a general all-of-medicine assistant. Its target domain is practical chest radiograph reading, especially settings in which a model must not only produce a final answer or report but also expose a clinically meaningful reasoning chain anchored to image-local evidence. The motivating claim is that prior open-source medical and radiology vision-LLMs can often produce useful answers or reports, yet typically do so as black boxes and are weak at tying reasoning to specific local regions (Lin et al., 4 Sep 2025).

The model is framed around two task families: full report generation and visual question answering. This scope matters because the authors explicitly connect chest X-ray interpretation to radiological workflow: radiologists inspect localized findings, relate them to anatomy and pathology, and synthesize findings and impressions. DeepMedix-R1 is intended to approximate that workflow by producing an answer together with intermediate reasoning steps grounded in local coordinates.

Within the paper’s own nomenclature, “R1” denotes a reasoning-oriented extension of a supervised DeepMedix baseline. The novelty is threefold: explicit reasoning traces, local grounding through image-region coordinates, and a sequential training pipeline that culminates in online reinforcement learning. This suggests that the system’s central contribution lies less in inventing a new base architecture than in shaping a pretrained multimodal model into a chest-X-ray-specific reasoner.

2. Model architecture and grounded output representation

Architecturally, DeepMedix-R1 is a specialization of Qwen2.5-VL-7B. The paper does not introduce a new low-level vision encoder or a separate localization head in the conventional detection sense. Instead, it adapts a multimodal autoregressive model so that the generated text itself carries localized grounding information. Inputs are a chest X-ray image $\mathcal{I}$ and a natural-language query or instruction $\mathcal{Q}$; outputs are structured sequences containing an intermediate reasoning process and a final answer (Lin et al., 4 Sep 2025).

The response format is central to the model’s identity. The format reward expects “the thinking process in the > ... tag and the final result in \boxed{} tag.” Grounding is generation-based rather than detector-based: the model writes coordinates for local image regions inside the reasoning trace, and these coordinates are rewarded if they are present and lie within valid image bounds. The paper therefore treats grounding as a property of structured language generation, not as the output of a separate supervised region-proposal module.

This design has two important implications. First, interpretability is surfaced directly in the textual output, since each query can elicit both a rationale and a final answer. Second, localization is expressed in a lightweight way that remains compatible with an autoregressive LLM. A plausible implication is that the model prioritizes integrated reasoning-and-grounding behavior over explicit detection-style modularity.

3. Sequential training pipeline

DeepMedix-R1 uses a three-stage sequential training pipeline. Stage 1 creates the supervised DeepMedix model, stage 2 provides cold-start grounded reasoning, and stage 3 applies online reinforcement learning to produce DeepMedix-R1 (Lin et al., 4 Sep 2025).

Stage	Data	Role
Stage 1	351.5K samples	CXR instruction fine-tuning
Stage 2	3.9K samples	Synthetic grounded reasoning initialization
Stage 3	6.6K samples	Online RL refinement

In stage 1, the authors curate a multi-source chest X-ray instruction dataset with 351.5K samples covering report generation, visual question answering, and localized image interpretation. This stage is intended to equip the model with broad foundational chest X-ray competence. The supervised objective is standard next-token prediction:

$$\mathcal{L}_{\rm SFT}=-\sum_{n=1}^N \log\big(x_n|x_{<n},\mathcal{I},\mathcal{O}\big).$$

The text states that the objective is based on the instruction $\mathcal{Q}$ and X-ray $\mathcal{I}$ and previously generated text tokens, while the displayed formula conditions on $\mathcal{I}$ and $\mathcal{O}$; the paper notes this as a notation inconsistency.

In stage 2, the model receives a cold-start reasoning initialization using 3.9K synthetic reasoning samples. These are produced by prompting GPT-4.1 with the X-ray, the task prompt, and the ground-truth answer, asking it to generate a reasoning process with grounded image representations. The generated reasoning procedures are then validated using Qwen2.5-VL-32B and filtered to remove clearly unreasonable examples, including cases where “the provided bounding box of the image region exceeds the maximum edge of the image.” The surviving samples are used for continued supervised fine-tuning with the same next-token objective.

In stage 3, the model is refined with online reinforcement learning on 6.6K samples. This stage is intended to improve both grounded reasoning quality and answer or report quality. The key design choice is that reward is assigned not only to final answer correctness but also to coordinate grounding and output formatting.

4. Reinforcement learning and grounded reasoning mechanism

The reinforcement-learning stage uses Group Relative Policy Optimization, or GRPO. The query distribution is written as $P(Q)$, each sampled query is $q=[\mathcal{Q},\mathcal{I}]$, the old policy is $\pi_{\theta_{\rm old}}$, the current policy is $\mathcal{Q}$0, the frozen reference policy is $\mathcal{Q}$1, and $\mathcal{Q}$2 is the number of responses sampled per question. The paper presents this as a PPO-style clipped surrogate objective with group-relative normalized advantages and a KL regularizer to the reference model, which in practice is the fine-tuned DeepMedix model (Lin et al., 4 Sep 2025).

The KL term is estimated with the unbiased estimator

$\mathcal{Q}$3

The paper’s typesetting for the full GRPO objective is imperfect, but its intended structure is explicit: multiple outputs are sampled for each question, rewards are computed over answer quality, format compliance, and coordinate grounding, and advantages are normalized within each sampled group.

Conceptually, the RL stage incentivizes a particular kind of reasoning behavior. The model is not only rewarded for being correct; it is also rewarded for being correct in the expected grounded format. This is the paper’s operational definition of “grounded reasoning.” Rather than adding a separate verifier or external tool router, DeepMedix-R1 makes the generation itself carry interpretable reasoning and spatial references.

5. Evaluation and reported performance

The paper reports substantial quantitative gains in both report generation and visual question answering. For report generation, DeepMedix-R1 improves by 14.54% and 31.32% over LLaVA-Rad and MedGemma. For visual question answering, it improves by 57.75% and 23.06% over MedGemma and CheXagent. To support evaluation of answer quality, the authors introduce Report Arena, a benchmarking framework that uses advanced LLMs to evaluate outputs (Lin et al., 4 Sep 2025).

Evaluation aspect	Comparator	Reported result
Report generation	LLaVA-Rad, MedGemma	14.54% and 31.32% improvement
Visual question answering	MedGemma, CheXagent	57.75% and 23.06% improvement
Expert review of reasoning	Qwen2.5-VL-7B	0.7416 vs. 0.2584 overall preference

The paper also reports an expert review focused on generated reasoning steps. In that comparison, DeepMedix-R1 is preferred to the established Qwen2.5-VL-7B model with an overall preference of 0.7416 versus 0.2584. The authors interpret this as evidence of greater interpretability and clinical plausibility.

These results are presented as evidence that explicitly trained grounded reasoning can improve both generation quality and the perceived quality of model explanations. A plausible implication is that the model’s gains are not limited to formatting compliance; they appear to extend to the substantive quality of chest-X-ray-specific reasoning as judged both quantitatively and by expert review.

6. Relation to adjacent systems, nomenclature, and recurrent points of confusion

DeepMedix-R1 belongs to a broader family of reasoning-oriented medical systems, but it is distinct from several similarly named or conceptually adjacent models. It differs from MedRAX, which is a chest-X-ray-specific ReAct-style agent that dynamically orchestrates external tools such as CheXagent, LLaVA-Med, MedSAM, Maira-2, report generation, DenseNet-121, and RoentGen without additional training; MedRAX externalizes reasoning through tool use and memory, whereas DeepMedix-R1 internalizes grounded reasoning within an end-to-end multimodal foundation model (Fallahpour et al., 4 Feb 2025).

It is also distinct from MediX-R1, an open-ended reinforcement learning framework for medical multimodal LLMs whose paper explicitly states that there is no mention anywhere of “DeepMedix-R1” as an alternate name, subproject, code name, or prior version (Mullappilly et al., 26 Feb 2026). Likewise, it is distinct from DeepMediX, a 2023 MobileNetV2-based resource-efficient classifier for brain MRI and skin lesion tasks that reports no “R1” variant in the paper (Nampalle et al., 2023).

Within the broader “R1” literature, DeepMedix-R1 sits alongside several domain-specific reasoning systems rather than standing alone as a universal template. MedReason-R1 applies GRPO and embedded local zoom to chest CT diagnosis; Med3D-R1 combines supervised fine-tuning and RL for 3D chest CT abnormality diagnosis; and RareDxR1 targets open-domain rare disease diagnosis from unstructured clinical notes through knowledge internalization, Reflection-Enhanced Reasoning Sampling, and dual-level curriculum reinforcement learning (Li et al., 22 Oct 2025, Lai et al., 1 Feb 2026, Jiang et al., 30 Jun 2026). This suggests that “R1” in current medical AI usage often denotes a training philosophy centered on explicit reasoning traces and reinforcement learning, but with strong task and modality specificity.

Several limitations or boundary conditions follow directly from the DeepMedix-R1 design. It is chest-X-ray-specific; it does not introduce a separate localization head in the conventional detection sense; and its grounding mechanism is generation-based, with coordinates emitted in text rather than produced by a standalone detector (Lin et al., 4 Sep 2025). The paper’s own training description also reveals dependence on synthetic reasoning initialization and on language-model-based evaluation in Report Arena. These do not negate the model’s contribution, but they define its methodological character: DeepMedix-R1 is best understood as an end-to-end, chest-radiograph-specific, RL-refined foundation model for grounded reasoning, not as a general medical agent, a detector-centric system, or a renamed version of DeepMediX or MediX-R1.