sandeep-shaw10/mendeley_knee_dxa

🦴 Knee X-ray Osteoporosis Dataset (Processed Derivative of Mendeley fxjm8fb6mw.2)

Original Source Dataset: Knee X-ray Osteoporosis Database
Published: 27 August 2021
Version: 2
DOI: 10.17632/fxjm8fb6mw.2
Contributors: Insha Majeed Wani, Sakshi Arora

This Hugging Face dataset is a processed and reformatted derivative of the publicly released Knee X-ray Osteoporosis Database on Mendeley Data. It combines knee X-ray images with tabular clinical, demographic, lifestyle, and bone health features for each participant, making it convenient for multi-modal machine learning research (image + structured data).

---

✅ What’s in This Hugging Face Version?

Component	Included	Notes
Knee X-ray images	✅	Mapped from filenames (N, OP, OS) to diagnosis classes.
Diagnosis Classes	✅	N=Normal, OP=Osteopenia, OS=Osteoporosis.
Patient Metadata	✅	Merged from the provided Excel sheet in the source dataset.
X-ray View Info	✅	Extracted from filename: OK=One Knee, TK=Both Knee.
Patient Code Join	✅	Filenames like OP1_TK.jpg matched to Excel Patient Id (e.g., OP1).
Image Preview in Hub	✅	image column cast to datasets.Image() so images display in the UI.

---

🧬 Dataset Summary (Original Description Adapted)

Osteoporosis is a major bone disorder (after arthritis) affecting millions globally. DXA (Dual Energy X-ray Absorptiometry) is the diagnostic gold standard but is expensive and not always accessible—particularly in resource-constrained regions. This database was created to support research into cost-effective, early, and scalable osteoporosis screening tools, including those that use knee X-rays as an imaging proxy along with clinical risk factors.

The clinical spreadsheet accompanying the X-rays includes:

• Demographics: age, gender, menopause age
• Anthropometrics: height, weight, BMI
• Medical factors: diabetes, hypothyroidism, seizure disorder, fracture history
• Lifestyle: smoking, alcohol, walking distance, eating habits
• Reproductive history: number of pregnancies, estrogen use
• Bone health measurements: T-score, Z-score (Quantitative Ultrasound System)
• Final diagnosis category (Normal / Osteopenia / Osteoporosis)

By combining these with knee imaging, the dataset supports data-centric research toward early, low-cost osteoporosis detection and risk stratification.

---

🔖 Labels & Filename Convention

Image filenames follow this pattern: _.jpg

Label Codes:

• N → Normal
• OP → Osteopenia
• OS → Osteoporosis

X-ray View Codes:

• OK → One Knee (single knee view)
• TK → Two Knee (bilateral view)

Examples:

Filename	Class	Patient ID	View
N19_TK.jpg	Normal	19	Two Knee
OP5_OK.png	Osteopenia	5	One Knee
OS12_TK.jpg	Osteoporosis	12	Both Knee

---

📁 Data Fields

Below are the fields in this processed dataset (names reflect normalized columns):

Core Image & Join Fields

• filename (string) – Image filename.
• image (image) – Knee X-ray (rendered inline on Hub).
• label (string) – Raw label code (N, OP, OS).
• patient_id (int64) – Numeric ID parsed from filename suffix.
• patient_code (string) – Full prefix from filename (e.g., OP5), used for Excel join.
• xray_code (string) – OK or TK.

Clinical Features (from Excel)

Below is a mapping of fields included from the source spreadsheet; names are lowercased / normalized and punctuation simplified.

Column	Type	Description
s.no	float64	Serial number in sheet.
joint_pain:	string	Reported joint pain.
gender	string	Gender.
age	float64	Age (yrs).
menopause_age	float64	Menopause age (yrs).
height__(meter)	float64	Height (m).
weight_(kg)	float64	Weight (kg).
smoker	string	Smoking status.
alcoholic	string	Alcohol use.
diabetic	string	Diabetes status.
hypothyroidism	string	Thyroid disorder.
number_of_pregnancies	float64	Gravidity.
seizer_disorder	string	Seizure disorder (as in spreadsheet).
estrogen_use	string	Estrogen usage.
occupation	string	Occupation category.
history_of_fracture	string	Prior fracture history.
dialysis:	string	Dialysis status.
family_history_of_osteoporosis	string	Family history.
maximum_walking_distance_(km)	float64	Mobility proxy.
daily_eating_habits	string	Nutrition pattern.
medical_history	string	Additional medical notes.
t-score_value	float64	Bone density T-score (QUS).
z-score_value	float64	Bone density Z-score (QUS).
bmi:	float64	Body Mass Index.
site	string	Measurement site.
obesity	string	Obesity category.
diagnosis	string	Ground-truth category (Normal / Osteopenia / Osteoporosis).

---

🔬 Recommended Use Cases

• Multimodal Classification: Train models that combine image + tabular features to predict diagnosis.
• Pre-screening Models: Develop cost-effective osteoporosis risk stratifiers when DXA is unavailable.
• Domain Adaptation: Compare performance across imaging vs. clinical-only vs. fusion models.
• Explainability Studies: Use tabular variables to interpret model predictions.

---

🚀 Quickstart

from datasets import load_dataset
from PIL import Image

# Replace with your actual dataset repo ID
ds = load_dataset("your-username/knee-xray-dataset")

# Inspect first record
example = ds["train"][0]
print(example.keys())

# Load image
img = example["image"]  # This is a PIL.Image.Image
img.show()

# Diagnosis
print("Diagnosis:", example["diagnosis"])
print("Label code:", example["label"], "X-ray view:", example["xray_code"])

📤 How This Dataset Was Built (Processing Pipeline)

• Download original Mendeley dataset (DOI: 10.17632/fxjm8fb6mw.2).
• Normalize filenames and parse components: label code, patient ID, X-ray view.
• Load patient metadata spreadsheet (“patient details.xlsx”).
• Normalize column names (lowercase, underscores).
• Join on patient_code (e.g., OP5).
• Export merged DataFrame and convert to Hugging Face Dataset.
• Cast image column to datasets.Image() so images render inline on Hub.
• Upload to Hugging Face Hub (this dataset).

📚 Citation If you use this dataset, please cite the original Mendeley dataset:

Knee X-ray Osteoporosis Database. Published 27 August 2021. Version 2. DOI: 10.17632/fxjm8fb6mw.2. Contributors: Insha Majeed Wani, Sakshi Arora.

BibTeX (generic form; adapt as needed):

@dataset{wani2021knee,
  title        = {Knee X-ray Osteoporosis Database},
  author       = {Wani, Insha Majeed and Arora, Sakshi},
  year         = {2021},
  publisher    = {Mendeley Data},
  version      = {2},
  doi          = {10.17632/fxjm8fb6mw.2},
  url          = {https://data.mendeley.com/datasets/fxjm8fb6mw/2}
}

---

🔐 Licensing & Usage

The original dataset is hosted on Mendeley Data. Please review the license terms on the source page: https://data.mendeley.com/datasets/fxjm8fb6mw/2

This processed version is shared for research and educational use. If license conflicts arise, the Hugging Face dataset may need to be made private or removed upon request.

---

🙏 Acknowledgments

• Original data curators: Insha Majeed Wani, Sakshi Arora, and collaborators.
• Screening data: Osteoporosis screening camp participants.
• Processing, formatting, and Hugging Face dataset assembly: [ Sandeep Shaw / IIT(ISM) ].

---

🧾 Changelog

Version	Date	Changes
2	August 27, 2021	- Added knee X-ray images for all participants.
		- Included clinical metadata: age, gender, menopause age, height, weight, and medical history.
		- Integrated T-score and Z-score from Quantitative Ultrasound System.
		- Improved formatting and consistency of metadata.
		- Enhanced dataset accessibility for global osteoporosis research.