---
title: 'WoundQwen: Multistage MLLM for Wound Diagnosis'
url: https://www.emergentmind.com/topics/woundqwen
type: topic
---

# WoundQwen: Multistage MLLM for Wound Diagnosis

WoundQwen is a three-stage multimodal large language model (MLLM) framework for surgical wound diagnosis and patient-facing wound care guidance introduced together with SurgWound-Bench, a benchmark built on the open SurgWound dataset. Its central design premise is that surgical wound assessment should not be reduced to direct image classification alone: the framework first predicts clinically salient wound characteristics, then uses those predictions as additional knowledge inputs for infection-risk and urgency assessment, and finally synthesizes a comprehensive report. In the originating formulation, WoundQwen is explicitly positioned against a prior landscape dominated by private datasets, image-only CNN classifiers, and systems that cannot incorporate richer clinical context such as wound location or produce understandable guidance [2508.15189].

## 1. Clinical task and conceptual scope

WoundQwen was proposed for **surgical wound diagnosis**, with emphasis on **infection risk**, **urgency guidance**, and **report generation** from surgical images plus structured wound attributes. The motivating clinical condition is surgical site infection (SSI), described as one of the most common and costly healthcare-associated infections; the paper frames surgical wound care as a setting in which earlier and more structured AI assistance could improve screening and downstream intervention decisions [2508.15189].

The framework separates two decision outputs that are clinically linked but not identical. **Infection Risk** is modeled with the labels **Low**, **Medium**, and **High**, whereas **Urgency Level** is modeled with **Home Care**, **Clinic Visit**, and **Emergency Care**. The paper notes that these outputs are strongly correlated, typically with low risk corresponding to home care, medium risk to clinic visit, and high risk to emergency care, while also noting that a small number of examples deviate slightly on surgeon judgment. This separation is important because it formalizes both etiologic assessment and actionable guidance rather than collapsing the task into a single label [2508.15189].

A common simplification is to treat WoundQwen as a monolithic image-to-text model. In the published formulation, it is instead a staged diagnostic system in which intermediate wound-characteristic predictions are retained and reused. This design makes the model closer to a structured reasoning pipeline than to a generic visual captioner or free-form medical chatbot.

## 2. SurgWound and SurgWound-Bench

WoundQwen is built on **SurgWound**, described as the first open-source dataset for diverse surgical wound analysis. The data were collected from publicly available social media and online platforms, specifically **RedNote, Twitter, Facebook, Instagram, and Reddit**, using domain hashtags, keywords, and posts from surgeons’ accounts. Filtering proceeded in two steps: **GPT-4o** first screened images for clear surgical wound content and quality, and then **three surgeons** manually verified the retained images and removed non-surgical or poor-quality cases [2508.15189].

The final dataset contains **686 images** in the main description, although the abstract reports **697**. The detailed dataset section and tables use **686**, split into **480 training**, **69 validation**, and **173 test** images in a **7:1:2** ratio. Each image is annotated by **three professional surgeons** with **8 structured clinical labels**:

- **Location**
- **Healing Status**
- **Closure Method**
- **Exudate Type**
- **Erythema Presence**
- **Edema Presence**
- **Infection Risk**
- **Urgency Level**

The label spaces are explicitly defined. **Location** includes **Abdomen, Ankle, Facial region, Manus, Patella, Cervical region, Other, Uncertain**. **Healing Status** includes **Healed** and **Not Healed**. **Closure Method** includes **Invisible, Staples, Sutures, Adhesives, Uncertain**. **Exudate Type** includes **Non-existent, Purulent, Serous, Sanguineous, Seropurulent, Uncertain**. **Erythema** and **Edema** each include **Non-existent, Existent, Uncertain**. **Infection Risk** includes **Low, Medium, High**, and **Urgency Level** includes **Home Care, Clinic Visit, Emergency Care**. The annotation scheme permits **Uncertain** when the value cannot be determined from the image alone, and **Other** is defined as a location that is identifiable but not among common categories [2508.15189].

The class distribution is highly imbalanced: most cases are **Low risk** and **Home Care**, and most wounds are **Healed** or have **No exudate**. The paper therefore analyzes statistical relationships between wound characteristics and infection risk using a chi-square test of independence, **Cramér’s V**, and **relative risk (RR)**. All tested attributes have \(p < 0.05\), indicating association with infection risk; the strongest associations are for **Healing Status**, **Exudate Type**, **Erythema**, and **Edema**, while **Location** and **Closure Method** are weaker. Using RR, the authors identify **Not Healed**, **Erythema**, **Edema**, and exudates such as **Serous, Sanguineous, Purulent, and Seropurulent** as elevated-risk factors [2508.15189].

On top of SurgWound, the authors define **SurgWound-Bench**, which contains two tasks. **SurgWound-VQA** is a multiple-choice benchmark derived from the annotations, excluding any attribute marked **Uncertain**; it evaluates seven sub-tasks: Healing Status, Closure Method, Exudate Type, Erythema, Edema, Infection Risk, and Urgency Level. **SurgWound-Report** uses test-set images with reference reports constructed through a two-step pipeline in which **GPT-4o** first generates a structured draft from the image and annotations, after which surgeons manually revise the draft for accuracy and consistency. VQA is evaluated with **Accuracy, Precision, Recall, F1, Micro-F1, Macro-F1, and Weighted-F1**, while report generation uses **BLEU-1, BLEU-2, BLEU-3, ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore** [2508.15189].

## 3. Three-stage architecture

The core contribution of WoundQwen is a staged MLLM architecture that decomposes diagnosis into wound-characteristic extraction, diagnostic outcome prediction, and report generation. The design is intended to mirror clinical reasoning: first identify local wound findings, then interpret their implications for risk and intervention, and only then compose narrative guidance [2508.15189].

| Stage | Inputs | Outputs |
|---|---|---|
| Stage 1 | Image | Healing Status, Closure Method, Exudate Type, Erythema, Edema |
| Stage 2 | Image, Location, Stage-1 predictions | Infection Risk, Urgency Level |
| Stage 3 | Image, Location, Stage-1 and Stage-2 outputs | Comprehensive wound report |

In **Stage 1**, WoundQwen trains **five separate Qwen-based MLLMs** to predict five wound characteristics from the image: **Healing Status**, **Closure Method**, **Exudate Type**, **Erythema**, and **Edema**. **Location is not predicted**; it is treated as known context. The predicted characteristic outputs are denoted as
$$
\{\hat{y}_i\}_{i=1}^{5}.
$$

In **Stage 2**, two additional models consume the image, the known **Location**, and the Stage-1 predictions. The paper formalizes the mechanism as
$$
\hat{y}_{\text{risk}} = \mathrm{WoundQwen\_risk}\bigl(\text{Image},\; \text{Location},\; \{\hat{y}_i\}_{i=1}^{5}\bigr)
$$
and
$$
\hat{y}_{\text{urgency}} = \mathrm{WoundQwen\_urgency}\bigl(\text{Image},\; \text{Location},\; \{\hat{y}_i\}_{i=1}^{5}\bigr).
$$
This is the framework’s defining “predictions as additional knowledge inputs” mechanism. Rather than discarding intermediate outputs, WoundQwen injects them as structured evidence for downstream reasoning [2508.15189].

In **Stage 3**, **WoundQwen_report** generates a **comprehensive wound report** conditioned on the raw image, wound characteristics, infection risk, urgency level, and known location. The paper describes this as a means of producing a clinically grounded narrative summary rather than a free-form hallucinated description. The resulting output is intended to resemble a clinical note or patient instruction more than a generic caption [2508.15189].

## 4. Model implementation and training regime

The Stage-1 models are based on **Qwen2.5-VL-7B**, initialized from **HuatuoGPT-Vision-7B** weights and fine-tuned with **LoRA supervised fine-tuning (SFT)**. The vision encoder is frozen, and LoRA adapters are applied to the LLM layers, specifically to the attention projections **\(q\_proj, k\_proj, v\_proj, o\_proj\)** and MLP projections **\(gate\_proj, up\_proj, down\_proj\)**. Training uses VQA-style instruction pairs such as “What is the healing status of this surgical wound?” [2508.15189].

Stage 2 also uses **LoRA-based SFT**, but with prompts augmented by wound-characteristic descriptions. The architectural point is not merely parameter-efficient fine-tuning; it is the explicit conditioning of risk and urgency prediction on a structured latent clinical state inferred in Stage 1. The authors report that these intermediate cues are clinically meaningful because wound healing status, exudate, erythema, and edema have statistically supported associations with infection risk in the dataset analysis [2508.15189].

Stage 3 is trained on the surgeon-refined image-report pairs from SurgWound-Bench. Its prompt concatenates the outputs from Stages 1 and 2 together with known location, creating a report-generation setting in which narrative synthesis is anchored by structured diagnostic context. This training design is explicitly motivated by reducing hallucination relative to direct image-to-text prompting [2508.15189].

A frequent misunderstanding is that the intermediate predictions are only auxiliary supervision. The paper’s formulation makes them central to the pipeline: they function as explicit, reusable clinical evidence. This suggests that WoundQwen should be interpreted as a structured multimodal reasoning system rather than a conventional single-pass classifier.

## 5. Quantitative performance and ablation evidence

Across the benchmark, WoundQwen is reported to outperform all compared baselines, including proprietary models **GPT-4o**, **Claude-3.5**, **Gemini-1.5**, **Gemini-2.0**, and open models such as **HuatuoGPT-34B**, **InternVL3-78B**, and **Qwen2.5-VL-72B**. The gains are reported for both the VQA tasks and the report-generation task, indicating that the framework improves not only categorical prediction but also report fidelity [2508.15189].

For **VQA**, the paper highlights several representative results. On **Exudate Type**, WoundQwen reaches **ACC 88.80%** and **weighted-F1 0.884**. On **Infection Risk**, it reaches **ACC 83.21%** and **weighted-F1 0.845**. On **Urgency Level**, it reaches **ACC 82.35%** and **weighted-F1 0.822**. The infection-risk classwise F1 values are **0.913** for **Low**, **0.522** for **Medium**, and **0.444** for **High**. The paper emphasizes that this is much stronger than the baselines on rare classes, many of which fail entirely on the high-risk class [2508.15189].

For **report generation**, WoundQwen obtains **BLEU-1 0.457**, **BLEU-2 0.300**, **BLEU-3 0.212**, **ROUGE-1 0.612**, **ROUGE-2 0.279**, **ROUGE-L 0.418**, and **BERTScore 0.917**. The next-best models are described as far behind on BLEU and ROUGE, indicating materially better n-gram overlap and content coverage, while WoundQwen also achieves the highest semantic similarity [2508.15189].

The ablation findings are as important as the headline scores. First, **HuatuoGPT-7B + LoRA SFT** improves over the base model but remains notably worse than the full three-stage framework. Second, feeding **ground-truth clinical cues** instead of Stage-1 predictions improves Stage-2 performance, confirming that the intermediate attributes are informative and that Stage-1 error propagates downstream. Third, even general models such as **GPT-4o** and **Claude-3.5** improve when given ground-truth wound cues, which indicates that the structured attributes themselves have diagnostic value independent of the particular backbone. The paper therefore interprets the performance gap between predicted cues and ground-truth cues as motivation for the staged design: iterative reasoning helps, but its effectiveness depends on accurate intermediate feature extraction [2508.15189].

## 6. Position within wound-AI research and principal limitations

WoundQwen belongs to a broader transition in wound AI from narrow image-only classifiers toward multimodal systems that integrate image evidence, structured clinical cues, and downstream communication tasks. Earlier wound models emphasized image classification, including multi-task CNN classification of five wound conditions from color images under real-world acquisition variability [2303.16522]. Subsequent multimodal systems showed that adding **location** metadata to image features can materially improve wound-type classification, as in an **Xception + GMRNN** architecture for diabetic, pressure, surgical, and venous ulcers [2505.08086]. Other work extends multimodal reasoning toward hospitalization-risk prediction by combining image-derived wound variables with clinician-filled variables in a two-stage pipeline using **Xception** and **LightGBM** [2602.09315]. Within wound-language tasks, a related but distinct direction is free-text wound care VQA for patient queries paired with one or two wound images, where retrieval and metadata guidance are used to generate concise clinical responses in the MEDIQA-WV 2025 setting [2511.10591].

Against that background, WoundQwen is distinguished by three features that co-occur in one framework: an **open surgical-wound dataset**, a **benchmark** spanning both VQA and report generation, and a **three-stage MLLM pipeline** whose central mechanism is the injection of structured intermediate predictions as knowledge inputs. This suggests that its main contribution is not only higher benchmark performance, but also a more explicit operationalization of staged clinical reasoning.

The paper is nonetheless explicit about constraints. The dataset is highly imbalanced; rare classes remain difficult. The abstract and main text report different dataset sizes (**697** versus **686**), which is a documentation inconsistency rather than a modeling claim. More substantively, the ablation results show that Stage-2 performance improves when ground-truth cues replace predicted ones, so the framework remains sensitive to upstream prediction quality. A second misconception is that the model is fully image-derived; in fact, **Location** is treated as known context rather than predicted. A plausible implication is that deployment quality will depend not only on image interpretation but also on the availability and correctness of structured metadata. The paper’s own summary of significance is therefore measured: WoundQwen is presented as a more interpretable and clinically useful wound-analysis system because it progresses from detailed wound-characteristic extraction to diagnostic judgment and finally to report generation, rather than because it resolves all uncertainties in automated wound assessment [2508.15189].

Source: https://www.emergentmind.com/topics/woundqwen