Dataset Card for SurgWound

SurgWound is the first open-source dataset for surgical wound analysis across multiple procedure types. SurgWound comprises 697 surgical wound images, each annotated by surgical experts at The Ohio State University Wexner Medical Center (OSWUMC). Each image is accompanied by high-quality labels covering six surgical wound characteristic attributes and two diagnostic outcomes attributes.

SurgWound-Bench is the first multimodal benchmark for surgical wound analysis, which includes two tasks: SurgWound-VQA and SurgWound-Report

Dataset Details

Dataset Description

SurgWound:

SurgWound comprises 697 surgical wound images, accompanied by high-quality labels covering six surgical wound characteristic attributes and two diagnostic outcomes attributes.

• Location (Abdomen, Patella, Ankle, Facial region, Manus, Cervical region, Other, Uncertain)
• Healing Status (Healed, Not Healed)
• Closure Method (Invisible, Sutures, Staples, Adhesives, Uncertain)
• Exudate Type (Non-existent, Serous, Sanguineous, Purulent, Seropurulent, Uncertain)
• Erythema (Non-existent, Existent, Uncertain)
• Edema (Non-existent, Existent, Uncertain)
• Infection Risk Assessment (Low, Medium, High)
• Urgency Level (Home Care (Green): Manage with routine care, Clinic Visit (Yellow): Requires professional evaluation within 48 hours, Emergency Care (Red): Seek immediate medical attention)

SurgWound-Bench

SurgWound is divided into training, validation, and test sets in a 7:1:2 ratio. The test set comprises 173 surgical wound images. Using these images and their corresponding annotations, we design two tasks:

• Visual Question Answering (VQA) Task
• Report Generation Task

Dataset Structure

Data Instances

The following is an example from SurgWound-Bench VQA Task:

"id": "76.jpg_closure_method",
"image_name": "76.jpg",
"field": "Closure Method",
"task_type": "multi_choice",
"image": base64encoded_image,
"question": "What is the closure method of this surgical wound?",
"options": [
    "Invisible",
    "Sutures",
    "Staples",
    "Adhesives",
    "Uncertain"
],
"answer": "Sutures"

The following is an example from SurgWound-Bench Report Generation Task:

"id": "606.jpg_report",
"image_name": "606.jpg",
"field": "Medical Report",
"task_type": "report_generation",
"image": base64encoded_image,
"question": "\nGiven a surgical wound image, generate a detailed medical report that includes the following aspects: wound location, wound status, closure method, exudate characteristics, presence of erythema, presence of edema, urgency level, infection risk assessment.\n",
"answer": "The wound assessment of the healed abdominal wound, closed with sutures, reveals normal imaging status with no signs of exudate, erythema, or edema. The infection risk assessment is low, indicating a favorable prognosis. The urgency level is classified as home care (Green), requiring routine care management. It is recommended to continue monitoring for any signs of infection or delayed healing, and to maintain good hygiene practices to prevent complications. If there are any concerns or changes in the wound condition, prompt follow-up with a healthcare provider is advised."

Data Split

SurgWound is divided into training, validation, and test sets in a 7:1:2 ratio.

• Train
  • VQA (3435)
  • Report Generation (480)
• Val
  • VQA (500)
  • Report Generation (69)
• Test
  • VQA (979)
  • Report Generation (137)

Dataset Creation

Source Data

We utilize a collection of domain-specific hashtags (e.g., #surgicalwoundinfection) and keywords (e.g., postoperative wound) to extract relevant content from various platforms, including RedNote, Twitter, Facebook, Instagram, and Reddit.

Data Collection and Processing

Data Collection

We collect surgical wound images from publicly available content on social media platforms. Specifically, we utilize a collection of domain-specific hashtags (e.g., #surgicalwoundinfection) and keywords (e.g., postoperative wound) to extract relevant content from various platforms, including RedNote, Twitter, Facebook, Instagram, and Reddit. In addition, we further expand the dataset by collecting images specifically from the social media accounts of surgeons and other medical professionals, where postoperative wound cases are often shared for educational or awareness purposes.

Data Filtering

We perform a two-stage filtering process involving both AI expert and human expert review to ensure that only high-quality images containing visible surgical wounds are included in the dataset. First, we leverage GPT-4o as an AI expert to automatically assess whether an image depicts a clear surgical wound, filtering out low-resolution images or those lacking any wound-related content. Subsequently, three surgeons serve as human experts to manually review the remaining images and exclude any that low-resolution or do not depict authentic surgical wounds.

Expert Annotation

To ensure high-quality annotation while optimizing expert effort, we estimate the difficulty level of each image based on the predicted Risk Level outputs from three MLLMs: GPT-4o, Claude 3.5, and Gemini 2. Images for which all three models consistently predict a Low risk level are considered low-difficulty, while those with inconsistent predictions or predicted as Medium or High risk are categorized as high-difficulty. For low-difficulty cases, a single surgeon is randomly assigned to perform the annotation. For high-difficulty cases, the image is independently annotated by two randomly assigned surgeons. If any disagreement arises between their annotations, a third surgeon is introduced to review both sets of annotations and make the final decision.