Clinical Wound Image & Description Dataset

A multi-modal medical imaging and clinical text dataset for automated wound classification, severity staging, and medical report generation.

---

📊 Dataset Overview

Metric	Value
Total Samples	1,564
Image Resolution	350 × 350 pixels (uniform)
Color Mode	RGB (3-channel)
Image Format	PNG
Dataset Size	~174.4 MB
Annotation File	metadata.csv
Number of Wound Types	10 classes (including Normal)
Severity Staging	3 levels (Mild, Moderate, Severe)
Avg. Description Length	~38.5 words (Range: 9 – 84 words)

---

📁 Directory Structure

Final/
├── README.md               # Dataset documentation
├── metadata.csv            # Ground truth annotations and text descriptions (1,564 entries)
└── images/                 # All 1,564 preprocessed PNG image files
    ├── abrasions_1.png
    ├── burns_1.png
    ├── diabetic_wounds_1.png
    ├── normal_1.png
    ├── pressure_wounds_1.png
    └── ...

---

📝 Annotation Schema (metadata.csv)

The dataset annotations are stored in metadata.csv with the following columns:

Column	Data Type	Description	Example
img_path	string	Relative filename of the image in images/	pressure_wounds_201.png
type	string	Diagnostic wound category (10 classes)	Pressure
severity	string	Clinical severity grading (Mild, Moderate, Severe)	Moderate
description	string	Detailed clinical text description of wound morphology	"The wound is a small, deep puncture site with surrounding erythematous skin..."

---

🏷️ Class & Severity Distributions

1. Wound Types (10 Classes)

Wound Type	Sample Count	Percentage
Pressure	335	21.4%
Diabetic	307	19.6%
Venous	245	15.7%
Surgical	215	13.7%
Bruises	135	8.6%
Burn	87	5.6%
Abrasion	84	5.4%
Laceration	70	4.5%
Normal	52	3.3%
Cut	34	2.2%
Total	1,564	100.0%

2. Severity Staging

Severity Level	Sample Count	Percentage
Severe	784	50.1%
Moderate	572	36.6%
Mild	208	13.3%
Total	1,564	100.0%

3. Type × Severity Cross-Tabulation

Wound Type	Mild	Moderate	Severe	Total
Abrasion	13	66	5	84
Bruises	87	48	0	135
Burn	12	27	48	87
Cut	13	19	2	34
Diabetic	8	83	216	307
Laceration	3	37	30	70
Normal	52	0	0	52
Pressure	7	119	209	335
Surgical	9	105	101	215
Venous	4	68	173	245
Total	208	572	784	1,564

---

🩺 Clinical Text Descriptions

Each sample includes an expert-level textual description of the wound morphology. Descriptions typically document:

• Dimensions & Depth: Approximate physical measurements and lesion depth.
• Wound Bed & Tissue: Presence of granulation, slough, fibrinous exudate, or necrosis.
• Peri-Wound Characteristics: Surrounding erythema, edema, discoloration, or maceration.
• Exudate & Drainage: Active bleeding, serous fluid, purulent discharge, or absence of drainage.
• Infection Assessment: Potential risks, visible signs of infection, or clean/healing indicators.

Sample Entries

• Pressure Wound (pressure_wounds_201.png, Moderate):

  "The wound is a small, deep puncture site with surrounding erythematous skin. There is no bleeding but there is a noticeable amount of fluid or pus around the wound. The wound edges are not well defined and there are visible signs of infection."

• Diabetic Wound (diabetic_wounds_280.png, Severe):

  "The wound is 3 cm by 2 cm in size with a depth of 1 cm. The tissue is granulating and the wound base is covered by fibrinous exudate. The edges are irregular and have a rolled appearance. The peri-wound skin is erythematous and warm to the touch. There is no odor or pain. The wound is clean and free of infection."

• Normal Skin (normal_14.png, Mild):

  "The skin appears healthy with no evidence of lesions or discoloration. There is no indication of infection or tissue damage."

---

🎯 Potential Applications

1. Computer Vision (Classification & Staging):

   • Multi-class wound type classification (10 classes).
   • Ordinal severity prediction (Mild $\rightarrow$ Moderate $\rightarrow$ Severe).
   • Joint multi-task learning (predicting Type and Severity concurrently).

2. Vision-Language Models (VLM / Multimodal LLMs):

   • Automated clinical report generation / wound image captioning.
   • Medical Visual Question Answering (VQA).
   • Fine-tuning multimodal models (e.g., LLaVA-Med, Med-Flamingo, BioMedCLIP).

3. Clinical Decision Support Systems (CDSS):

   • Triage assistance and telemedicine evaluation for chronic wound management (diabetic ulcers, pressure ulcers, venous ulcers).

---

💻 Quickstart Loading Example (Python)

import os
import pandas as pd
from PIL import Image

DATASET_DIR = "./Final"
csv_path = os.path.join(DATASET_DIR, "metadata.csv")
images_dir = os.path.join(DATASET_DIR, "images")

# Load annotations
df = pd.read_csv(csv_path)
print(f"Loaded {len(df)} samples.")
print(df.head())

# Load first image
sample = df.iloc[0]
image_path = os.path.join(images_dir, sample["img_path"])
image = Image.open(image_path)
print(f"Image shape: {image.size}, Mode: {image.mode}")
print(f"Type: {sample['type']} | Severity: {sample['severity']}")
print(f"Description: {sample['description']}")