TLDR: Researchers have developed SurgWound-Bench, the first open-source dataset and benchmark for diverse surgical wound diagnosis, addressing critical gaps in data privacy and expert annotation costs. Alongside this, they introduced WoundQwen, a novel three-stage AI framework that uses multimodal large language models to analyze wound characteristics, predict infection risk and urgency, and generate comprehensive reports, significantly outperforming existing AI models.
Surgical site infections (SSIs) represent a significant challenge in healthcare, leading to increased patient suffering and substantial costs. Effective surgical wound care is crucial for preventing these infections and improving patient recovery. While artificial intelligence (AI) has shown promise in preliminary wound screening, progress has been hampered by concerns over data privacy, the high cost of expert medical annotation, and a lack of publicly available, diverse datasets specifically for surgical wounds.
Addressing these critical gaps, a team of researchers from The Ohio State University, The Ohio State University Wexner Medical Center, and Northeastern University has introduced a groundbreaking initiative: SurgWound-Bench: A Benchmark for Surgical Wound Diagnosis. This work presents the first open-source dataset and a comprehensive benchmark designed to advance AI capabilities in surgical wound assessment.
The SurgWound Dataset: A Foundation for AI Innovation
At the heart of this research is SurgWound, the first open-source dataset featuring a wide array of surgical wound types. It comprises 697 real-world surgical wound images, meticulously collected from various social media platforms using specific hashtags and keywords to ensure a diverse and representative sample. To guarantee data quality and ethical considerations, the images underwent a rigorous two-stage filtering process involving both AI (GPT-4o) and human expert surgeons, removing low-resolution or non-surgical wound content.
Each image in the SurgWound dataset is richly annotated by three professional surgeons with eight fine-grained clinical attributes. These attributes include wound location, healing status, closure method, exudate type, presence of erythema (redness), presence of edema (swelling), infection risk assessment, and urgency level for treatment. This detailed annotation provides a robust foundation for training and evaluating sophisticated AI models.
SurgWound-Bench: A New Standard for Evaluation
Building upon the SurgWound dataset, the researchers established SurgWound-Bench, the first multimodal benchmark for surgical wound analysis. This benchmark includes two primary tasks: Visual Question Answering (VQA) and Report Generation. The VQA task evaluates an AI model’s ability to accurately extract specific clinical attributes from wound images in a question-and-answer format. The report generation task challenges models to produce coherent and informative textual summaries of the wound condition, mimicking the detailed reports clinicians generate.
To ensure a thorough evaluation, the benchmark employs a comprehensive set of metrics, including accuracy, precision, recall, and various F1-scores for VQA, and BLEU, ROUGE, and BERTScore for report generation. These metrics are particularly important for medical data, which often suffers from class imbalance, ensuring that models perform well across all categories, including rare but critical conditions.
WoundQwen: A Three-Stage AI Diagnostic Framework
To leverage the new dataset and benchmark, the team also proposed WoundQwen, an innovative three-stage learning framework based on multimodal large language models (MLLMs). This framework is designed to provide personalized wound care instructions and timely interventions:
Stage 1: Surgical Wound Characteristic Analysis
Five independent MLLMs are trained to predict specific wound characteristics like healing status, closure method, exudate type, erythema, and edema from the images.
Stage 2: Surgical Wound Outcome Prediction
The predictions from Stage 1, along with the known wound location, serve as additional knowledge inputs to two MLLMs. These models, WoundQwen_risk and WoundQwen_urgency, are responsible for diagnosing infection risk and guiding subsequent interventions by assessing the urgency level.
Stage 3: Surgical Wound Report Generation
A final MLLM, WoundQwen_report, integrates all the diagnostic results from the previous two stages to produce a comprehensive and clinically relevant report. This stage aims to provide detailed analysis and subsequent instructions to patients.
The experimental results demonstrate that WoundQwen consistently outperforms current state-of-the-art MLLMs, including proprietary API models like GPT-4o, Claude-3.5, and Gemini, across all VQA sub-tasks and report generation metrics. This superior performance highlights WoundQwen’s powerful multimodal reasoning capabilities and the effectiveness of its domain-specific fine-tuning.
Also Read:
- LLM4Sweat: A New AI Framework for Hyperhidrosis Care
- AI Advances Fracture Detection in Wrist and Hand X-rays
Paving the Way for Future Wound Care
The introduction of SurgWound, SurgWound-Bench, and the WoundQwen framework marks a significant advancement in the field of AI-based surgical wound diagnosis. By providing the first open-source dataset and benchmark, along with a state-of-the-art diagnostic tool, this research lays a crucial foundation for future developments in personalized wound care, enabling earlier detection of infections, reducing clinic visits, and ultimately improving patient outcomes. The dataset and code are openly released to support further research and development in this critical area. You can find more details about this research paper here.
