ayushirathour/chest-xray-pneumonia-detection

PneumoDetectAI

Binary classification model for pneumonia detection in pediatric chest X-rays (ages 1-5). Built with TensorFlow and MobileNetV2, validated on independent operator cohort with 86% accuracy and 96.4% sensitivity.

Author: Ayushi Rathour
Contact: ayushirathour1804@gmail.com
Framework: TensorFlow 2.19
Model Size: ~14 MB

Model Overview

PneumoDetectAI is a deep learning model designed to detect pneumonia in chest X-rays of pediatric patients aged 1 to 5 years. The model uses transfer learning from MobileNetV2 for efficient inference while maintaining clinically relevant performance.

Key Specifications

Property	Value
Architecture	MobileNetV2 (ImageNet pretrained) + custom head
Input Shape	224 × 224 × 3 (RGB)
Output	Binary classification (NORMAL/PNEUMONIA)
File Format	TensorFlow SavedModel (.h5)
Model Size	~14 MB
Inference Time	0.46 seconds on CPU
Target Population	Pediatric patients (1-5 years)

Intended Users

• ML researchers working on medical imaging
• Healthcare AI developers building screening tools
• Students learning medical AI validation approaches
• Radiologists interested in AI-assisted screening

Performance Metrics

Validation Type	Dataset	Samples	Accuracy	Sensitivity	Specificity
Internal	Mooney 2018	269	94.8%	89.6%	100%
Cross-Operator	Radiography 2024	485	86.0%	96.4%	74.8%

Clinical Interpretation

• High Sensitivity (96.4%): Catches 96 out of 100 pneumonia cases, suitable for screening
• Moderate Specificity (74.8%): 25% false positive rate acceptable for screening tool
• Generalization: 8.8% accuracy drop on independent cohort indicates reasonable robustness

Quick Start Usage

from huggingface_hub import hf_hub_download
import tensorflow as tf
import numpy as np
from PIL import Image

# Download and load model
model_path = hf_hub_download(
    repo_id="ayushirathour/chest-xray-pneumonia-detection",
    filename="best_chest_xray_model.h5"
)
model = tf.keras.models.load_model(model_path)

# Preprocess image
def preprocess_xray(image_path):
    img = Image.open(image_path).convert("RGB").resize((224, 224))
    img_array = np.array(img) / 255.0
    return np.expand_dims(img_array, axis=0)

# Make prediction
image_array = preprocess_xray("chest_xray.jpg")
probability = model.predict(image_array)[0][0]
diagnosis = "PNEUMONIA" if probability >= 0.5 else "NORMAL"
confidence = probability * 100 if probability >= 0.5 else (1 - probability) * 100

print(f"Diagnosis: {diagnosis}")
print(f"Confidence: {confidence:.1f}%")

Training Details

Datasets

• Training Data: Chest X-Ray Images (Pneumonia) by Paul Timothy Mooney

  • Source: Guangzhou Women and Children's Medical Center
  • Size: ~5,863 images (pediatric patients aged 1-5)
  • Split: Pre-divided train/validation/test

• External Validation: Pneumonia Radiography Dataset by Tanmay Shukla

  • Source: Same hospital, different operators and time period
  • Size: 485 independent samples
  • Purpose: Cross-operator generalization testing

Architecture Details

• Base Model: MobileNetV2 (ImageNet weights frozen initially)
• Custom Head: Global Average Pooling → Dropout (0.5) → Dense (128) → Dense (1, sigmoid)
• Optimization: Adam optimizer (lr=0.0001)
• Loss Function: Binary crossentropy
• Training: 20 epochs with early stopping

Limitations & Risks

Technical Limitations

• Single Institution: Both datasets from same medical center
• Age Restriction: Validated only on pediatric patients (1-5 years)
• Binary Output: Cannot distinguish pneumonia subtypes (viral vs bacterial)
• Image Quality: Performance degrades with poor quality or non-standard views

Clinical Limitations

• False Positive Rate: 25.2% may increase radiologist workload
• Screening Only: Not suitable for definitive diagnosis
• Population Bias: Trained on Asian pediatric cohort only
• No Clinical Context: Cannot incorporate patient history or symptoms

Deployment Risks

• Overconfidence: High sensitivity may create false sense of security
• Misuse: Risk of use without proper medical oversight
• Generalization: Performance may vary on different imaging equipment

Responsible AI & Ethics

Bias Considerations

• Population Bias: Model trained exclusively on Asian pediatric population
• Institutional Bias: Single medical center may not represent global imaging practices
• Age Bias: Performance on other age groups unknown

Required Safeguards

• Human Oversight: All predictions must be reviewed by qualified radiologists
• Screening Context: Should only be used as preliminary screening tool
• Informed Consent: Patients must be informed of AI involvement in screening
• Quality Assurance: Regular monitoring of real-world performance required

Regulatory Status

• Not FDA Approved: Research prototype only
• Not CE Marked: Not approved for clinical use in EU
• Research Use: Intended for academic and development purposes only

Citation

@misc{rathour2025pneumodetectai,
    title={PneumoDetectAI: Pediatric Chest X-Ray Pneumonia Detection with Cross-Operator Validation},
    author={Rathour, Ayushi},
    year={2025},
    note={Cross-operator validation on 485 independent samples},
    url={https://huggingface.co/ayushirathour/chest-xray-pneumonia-detection}
}

Acknowledgements

Datasets

• Training Dataset: Chest X-Ray Images (Pneumonia) - Paul Timothy Mooney (Kaggle)
• Validation Dataset: Pneumonia Radiography Dataset - Tanmay Shukla (Kaggle)
• Original Research: Kermany et al., "Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning", Cell 2018

Technical Stack

• Framework: TensorFlow 2.19
• Architecture: MobileNetV2 (Google)
• Deployment: Streamlit, FastAPI
• Hosting: Hugging Face Hub

Additional Resources

• 🌐 Live Demo: PneumoDetectAI Web App
• 📂 Source Code: GitHub Repository
• 📖 API Documentation: Available when running locally
• 💬 Issues & Support: GitHub Issues or email contact

---

Disclaimer: This model is for research and educational purposes only. It is not a medical device and should not be used for clinical diagnosis without appropriate medical supervision.