SurgWound: Surgical Wound Benchmark
- SurgWound is a public diagnostic benchmark and dataset for surgical wounds, encompassing diverse wound types and eight clinically relevant attributes.
- The dataset is curated from social media and annotated through a rigorous two-stage filtering process by expert surgeons, ensuring high-quality multimodal data.
- The benchmark supports visual question answering and report generation tasks, demonstrating modular, interpretable clinical reasoning with strong empirical performance.
SurgWound refers most specifically to the public surgical-wound resource introduced in "SurgWound-Bench: A Benchmark for Surgical Wound Diagnosis", where surgical wound analysis is formalized as a multimodal diagnostic problem spanning fine-grained attribute recognition, infection-risk assessment, urgency triage, and report generation. The resource was introduced against the backdrop that surgical site infection is one of the most common and costly healthcare-associated infections, while no publicly available dataset or benchmark encompassed various types of surgical wounds; accordingly, the work contributes the first open-source dataset featuring a diverse array of surgical wound types and the first benchmark for surgical wound diagnosis (Xu et al., 21 Aug 2025).
1. Scope and terminology
In the recent wound-AI literature, the term "SurgWound" appears in two closely related senses. First, it denotes the surgical-wound category used in multi-class wound classification settings, such as the four-category system of diabetic, pressure, surgical, and venous ulcers (Patel et al., 2023) and the five-class setting of burn injury, pressure injury, diabetic foot ulcer, vascular ulcer, and surgical wound (Tusar et al., 28 Mar 2026). Second, and more specifically, it denotes the dataset underlying SurgWound-Bench, which is centered exclusively on surgical wound diagnosis and is annotated with eight fine-grained clinical attributes (Xu et al., 21 Aug 2025).
This distinction is important because earlier systems for postoperative wound analysis did not define a public surgical-wound benchmark in the same sense. Deepwound, for example, is a multi-label convolutional neural network ensemble for postoperative wound assessment that identifies nine labels, including surgical site infection, drainage, fibrinous exudate, granulation tissue, open wound, staples, steri strips, and sutures, and is paired with a mobile application frontend (Shenoy et al., 2018). SurgWound-Bench shifts the emphasis from postoperative image tagging alone toward structured diagnosis and report synthesis (Xu et al., 21 Aug 2025).
A plausible implication is that SurgWound marks a transition from wound-image screening tasks to benchmarked clinical reasoning tasks over surgical images. That inference is consistent with the benchmark’s inclusion of visual question answering and report generation rather than only segmentation or categorical image classification (Xu et al., 21 Aug 2025).
2. Dataset constitution and annotation schema
The SurgWound dataset contains 686 surgical wound images in the analyzed set, although the abstract reports 697; the documented split is 480 training, 69 validation, and 173 test images, corresponding to a 7:1:2 partition (Xu et al., 21 Aug 2025). Images were sourced openly from social media and online platforms, including Twitter, Facebook, Instagram, Reddit, RedNote, and accounts of medical professionals, using surgical wound-related hashtags and keywords (Xu et al., 21 Aug 2025). The dataset is intended to capture diversity in wound types, locations, closure methods, and clinical presentations (Xu et al., 21 Aug 2025).
Annotation was performed by three professional surgeons from the Ohio State University Wexner Medical Center (Xu et al., 21 Aug 2025). The curation protocol used a two-stage filtering process: GPT-4o for initial filtering and three surgeons for manual quality control (Xu et al., 21 Aug 2025). Images with high annotation difficulty, identified by disagreement among three MLLMs or by risk above "Low", were double-labeled, and disputes were resolved by a third surgeon (Xu et al., 21 Aug 2025). This produces a label structure oriented toward both descriptive morphology and downstream management.
The eight annotated attributes are location, healing status, closure method, exudate type, erythema presence, edema presence, infection risk, and urgency level (Xu et al., 21 Aug 2025). The main documented value sets are as follows.
| Attribute | Documented values |
|---|---|
| Location | Abdomen, Ankle, Facial region, Manus, Patella, Cervical region, Other, Uncertain |
| Healing status | Healed, Not Healed |
| Closure method | Invisible, Staples, Sutures, Adhesives, Uncertain |
| Exudate type | Non-existent, Purulent, Serous, Sanguineous, Seropurulent, Uncertain |
| Erythema presence | Existent, Non-existent, Uncertain |
| Edema presence | Existent, Non-existent, Uncertain |
| Infection risk | Low, Medium, High |
| Urgency level | Home Care, Clinic Visit, Emergency Care |
The reported class distribution is strongly imbalanced. Examples given in the benchmark include Healed at 59.2% versus Not Healed at 40.8%, Sutures at 51%, Non-existent exudate at 77.4%, Non-existent erythema at 68.4%, Non-existent edema at 69.0%, Home Care at 87.5%, and Low infection risk at 83.5% (Xu et al., 21 Aug 2025). For report generation, reference reports were first produced by GPT-4o conditioned on annotations and then reviewed and edited by surgeons to guarantee clinical accuracy (Xu et al., 21 Aug 2025).
3. Benchmark tasks and evaluation protocol
SurgWound-Bench defines two tasks: visual question answering and medical report generation (Xu et al., 21 Aug 2025). In the VQA setting, each image is paired with attribute-specific questions, excluding cases labeled "Uncertain", so that the model predicts the clinically relevant attribute value from the image (Xu et al., 21 Aug 2025). In the report-generation setting, the goal is to generate a clinically coherent written report summarizing wound condition and recommendations (Xu et al., 21 Aug 2025).
The VQA evaluation uses accuracy, precision, recall, F1-score, Micro-F1, Macro-F1, and Weighted-F1 to address severe class imbalance (Xu et al., 21 Aug 2025). Weighted-F1 is defined as
where is the number of samples in class (Xu et al., 21 Aug 2025). The report-generation task is evaluated with BLEU-1, BLEU-2, ROUGE-1, ROUGE-L, and BERTScore (Xu et al., 21 Aug 2025).
This task design differentiates SurgWound-Bench from segmentation-centered wound benchmarks. WoundAmbit, for instance, standardizes training and evaluation for twelve semantic segmentation models, studies out-of-distribution generalization, computational efficiency, and interpretability, and converts AI-generated masks into wound size estimates (Borst et al., 8 Apr 2025). By contrast, SurgWound-Bench is organized around attribute-level diagnosis and narrative output (Xu et al., 21 Aug 2025).
4. WoundQwen and the three-stage diagnostic formulation
The reference framework for SurgWound-Bench is WoundQwen, a three-stage hierarchical and modular system built with multimodal LLMs (Xu et al., 21 Aug 2025). In Stage 1, five independent MLLMs predict Healing Status, Closure Method, Exudate Type, Erythema, and Edema (Xu et al., 21 Aug 2025). These models are based on Qwen2.5-VL-7B, initialized from HuatuoGPT-Vision-7B, trained with LoRA-based supervised fine-tuning, and use frozen vision encoders (Xu et al., 21 Aug 2025).
In Stage 2, two Qwen-based models, WoundQwen_risk and WoundQwen_urgency, diagnose infection risk and urgency level using the image, location, and all Stage 1 predictions as structured cues (Xu et al., 21 Aug 2025). The formulation is given as
Stage 3 trains WoundQwen_report, which takes the image together with predicted characteristics and outcomes to generate a complete report (Xu et al., 21 Aug 2025).
The framework is explicitly modular and interpretable: stage-wise predictions can be inspected and validated by clinicians, in contrast to black-box end-to-end models (Xu et al., 21 Aug 2025). At the same time, the benchmark notes a structural limitation: errors in Stage 1 can propagate to Stage 2 and Stage 3 (Xu et al., 21 Aug 2025). This makes WoundQwen a benchmark not only of recognition quality but of compositional diagnostic reliability.
5. Empirical performance and benchmark significance
On the VQA task, WoundQwen is reported to outperform GPT-4o, Claude-3.5, Gemini, HuatuoGPT-34B, Qwen2.5-VL-72B, and InternVL3-78B on all seven sub-tasks (Xu et al., 21 Aug 2025). For infection risk, the benchmark reports Accuracy 83.21% and Weighted-F1 0.845 for WoundQwen, compared with Accuracy 67.15% and Weighted-F1 0.727 for the best prior baseline, HuatuoGPT-34B (Xu et al., 21 Aug 2025). The paper attributes particular strength to WoundQwen on rare and imbalanced categories, including exudate types (Xu et al., 21 Aug 2025).
On report generation, WoundQwen achieves BLEU-1 0.457, BLEU-2 0.300, ROUGE-1 0.612, ROUGE-L 0.418, and BERTScore 0.917, surpassing all listed baselines (Xu et al., 21 Aug 2025). General-purpose models are described as exhibiting more hallucination and producing less clinically accurate text (Xu et al., 21 Aug 2025). Ablation studies further show that WoundQwen_risk and WoundQwen_urgency improve to accuracy up to approximately 88% when provided with ground-truth characteristics instead of Stage 1 predictions, indicating that better first-stage attribute recognition would further improve downstream diagnosis (Xu et al., 21 Aug 2025).
Several limitations are documented. The dataset is relatively small, the public-image source may introduce selection bias or incompletely capture some wound types, and the hierarchical structure allows module-level error propagation (Xu et al., 21 Aug 2025). The discrepancy between 697 images in the abstract and 686 in the analyzed split is a notable data-accounting detail rather than a conceptual inconsistency (Xu et al., 21 Aug 2025).
6. Position within the broader literature on surgical wound AI
SurgWound-Bench sits at the intersection of several previously separate lines of work. One line centers on postoperative wound image assessment. Deepwound used a modified VGG-16-based ensemble trained on 1,335 anonymized wound images and achieved multi-label classification over nine postoperative findings while supporting remote tracking through a proof-of-concept iOS app called Theia (Shenoy et al., 2018). Another line focuses on wound-type classification. The image-and-location model of "Integrated Image and Location Analysis for Wound Classification" used wound images plus a 484-region body map to classify diabetic, pressure, surgical, and venous ulcers; in ROI classification with location, the six-class setting reached 87.50% accuracy in one split, and binary classification for normal skin versus surgical wound reached 97.56–98.50% (Patel et al., 2023).
A third line emphasizes joint boundary segmentation and wound-type recognition. The YOLOv11-based study on joint wound boundary segmentation and multi-class wound classification included surgical wound as one of five classes and reported, for YOLOv11x, F1-scores of 0.9341 for wound boundary segmentation and 0.8736 for wound classification, with surgical wound achieving AUROC 1.00 in the one-vs-rest analysis (Tusar et al., 28 Mar 2026). In that literature, surgical wound is one class among several wound types. SurgWound-Bench instead isolates surgical wounds and expands the task to clinically structured reasoning (Xu et al., 21 Aug 2025).
A fourth line emphasizes telemedical deployment. WoundAIssist provides a patient-centered mobile app with on-device wound segmentation, physician oversight, questionnaires, messaging, and video consultations (Borst et al., 6 Jun 2025). WoundAmbit integrates wound segmentation, reference-object-based size retrieval, generalization analysis, and telehealth deployment (Borst et al., 8 Apr 2025). DM-WAT combines smartphone wound images with clinical notes from EHRs for referral decisions, using DeiT-Base-Distilled for images, DeBERTa-base for text, and intermediate fusion, achieving 77% with std 3% accuracy and 70% with std 2% F1 score (Fard et al., 22 Jan 2025). This suggests that SurgWound-Bench is best understood not as a replacement for segmentation, referral, or remote-monitoring systems, but as a benchmark layer for surgical-wound diagnosis that could, in principle, complement those systems.
From a clinical-informatics perspective, SurgWound also aligns with earlier work on dynamic postoperative surveillance. Prognostic modeling of surgical site infection using daily wound and clinical variables showed that dynamic wound data can support prediction of if and when SSI will occur, rather than static yes-or-no screening alone (Ke et al., 2016). A plausible implication is that future surgical-wound systems may combine SurgWound-style image-grounded attribute reasoning with longitudinal and multimodal monitoring, although that integration is not itself established by the present benchmark.