Abstract
This research presents an innovative approach to pediatric chest pneumonia classification using deep learning techniques combined with Generative Adversarial Networks (GANs) for addressing class imbalance in medical imaging datasets. We address the critical challenge of limited and imbalanced medical data by employing GANs for synthetic data generation, combined with traditional data augmentation methods. Our methodology demonstrates significant improvements in classification accuracy and model robustness for pneumonia detection in pediatric chest X-rays.
Through systematic experiments across four different data balancing strategies, we demonstrate that the combined approach (augmentation + GAN generation) achieves 87.18% accuracy, outperforming baseline methods and establishing a new benchmark for pediatric pneumonia classification.
 }})
Interactive Web Application: Real-time chest X-ray classification with probability scores
Problem Statement
Clinical Challenge: Pneumonia is a leading cause of mortality in children under 5 years globally. Early and accurate diagnosis through chest X-ray analysis is crucial for effective treatment.
Technical Challenge: Medical imaging datasets often suffer from class imbalance, where normal cases significantly outnumber pneumonia cases, leading to biased model performance.
Our Solution: We address these challenges using a novel combination of CNN architecture with GAN-based synthetic data generation and traditional augmentation techniques.
Methodology
Our Approach:
1. Data Preprocessing
- Image resizing to 148×148 pixels
- Grayscale conversion
- Pixel normalization [0,1]
2. GAN-based Generation
- DCGAN architecture
- 40,000 training iterations
- 2,534 synthetic normal X-rays
3. Data Augmentation
- Rotation (±35°)
- Width/Height shifts
- Shear & Zoom transforms
4. CNN Classification
- Optimized architecture
- Batch normalization
- Binary classification
Experimental Results
Key Finding: Our combined approach (traditional augmentation + GAN generation) achieved the highest performance with 87.18% accuracy, demonstrating the effectiveness of hybrid data balancing strategies in medical imaging.
| Experiment | Data Strategy | Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|---|
| Baseline | Original Imbalanced | 85.26% | 0.85 | 0.85 | 0.85 |
| Augmentation | Traditional Augmentation | 86.06% | 0.86 | 0.86 | 0.86 |
| GAN Generation | Synthetic Data Only | 84.78% | 0.85 | 0.85 | 0.85 |
| Combined | Augmentation + GAN | 87.18% | 0.87 | 0.87 | 0.87 |
Key Insights:
- Data Balancing Effectiveness: Combined approach achieved highest accuracy
- GAN Quality: Generated images maintained realistic chest X-ray characteristics
- Model Robustness: Balanced datasets showed improved generalization
- Clinical Relevance: High recall for pneumonia detection (87%)
Model Architecture
CNN Architecture:
model = Sequential([
Conv2D(32, (3, 3), input_shape=(148, 148, 1), activation='relu'),
MaxPooling2D(pool_size=(2, 2)),
BatchNormalization(),
Conv2D(64, (3, 3), activation='relu'),
MaxPooling2D(pool_size=(2, 2)),
Flatten(),
Dense(1, activation='sigmoid')
])
# Hyperparameters:
# - Optimizer: Adam
# - Loss: Binary Crossentropy
# - Image Size: 148×148
# - Batch Size: 132
# - Epochs: 15-50
🧠 Generator Network
4-layer transposed convolution network for synthetic X-ray generation
🔍 Discriminator Network
4-layer convolution network for authentic vs synthetic classification
⚡ Training Process
40,000 iterations with RMSprop optimizer and adversarial loss
📊 Output Quality
High-quality 148×148 synthetic chest X-ray images
🩺 Try the Model
Interactive Pneumonia Detection Demo
Upload a pediatric chest X-ray image directly to our live model hosted on Hugging Face Spaces
Click the button above to access the interactive model demo
📝 Note: If the embedded demo above doesn't load properly, please use the "Open Live Demo in New Tab" button to access the full application.
💡 How to Use:
- Upload Image: Click the upload area and select a chest X-ray image
- Analyze: The model will automatically process your image
- View Results: Get instant classification results with probability scores
- Medical Note: Results are for research purposes only - consult medical professionals for diagnosis
📊 Model Performance
Accuracy: 87.18%
Precision: 0.87
Recall: 0.87
⚙️ Technical Specs
Input Size: 148×148 pixels
Architecture: CNN + GAN
Framework: TensorFlow/Keras
Dataset Information
📊 Dataset Statistics:
Total Images
5,863 chest X-rays
Training Set
5,216 images
Test Set
624 images
Classes
Normal vs Pneumonia
Source: Kaggle Chest X-Ray Pneumonia Dataset
Class Distribution: The dataset exhibits class imbalance with more pneumonia cases than normal cases, making it an ideal testbed for our GAN-based balancing approach.
Future Directions
🔬 Multi-class Classification
Extend to different types of pneumonia and lung conditions
🏗️ Advanced Architectures
Implement ResNet, DenseNet, or Vision Transformers
🔐 Federated Learning
Enable privacy-preserving collaborative training
📱 Mobile Deployment
Optimize for real-time mobile medical applications
🏥 Clinical Validation
Collaborate with medical institutions for validation
🤖 Explainable AI
Implement attention mechanisms for decision transparency
Project Information
📧 Contact
Authors: Abdul Manaf, Nimra Mughal
Institution: Department of Computer Science
Research Area: Medical AI & Deep Learning
Status: Under Review
This research contributes to the advancement of AI-assisted medical diagnosis, specifically addressing the critical need for accurate pediatric pneumonia detection.