Deep Learning Project for Pneumonia Detection in X-ray Images Using Convolutional Neural Networks

This project is an open-source pneumonia detection project based on Convolutional Neural Networks (CNN), aiming to automatically classify chest X-ray images into normal or pneumonia categories. It supports both synthetic datasets and real Kaggle datasets, and compares the performance differences between Adam and SGD optimizers, providing practical references for medical AI-assisted diagnosis. The original author is rafayraza-nextgen, and the project is open-sourced on GitHub (link: https://github.com/rafayraza-nextgen/Pneumonia-Detector), released on June 13, 2026.

Pneumonia is one of the leading causes of death globally, and timely diagnosis is crucial, especially in areas with limited medical resources. Traditional X-ray diagnosis relies on the experience of radiologists; when there is a shortage of doctors, AI-assisted systems can improve screening efficiency. This project uses CNN to achieve automatic classification of X-ray images, helping to quickly screen suspected cases, and is suitable for telemedicine and primary care scenarios.

Core Network Structure

• Contains 3 convolutional layers (32, 64, 128 filters) to extract features layer by layer
• Batch Normalization is added after convolutional layers to accelerate convergence
• MaxPooling layer to reduce dimensionality
• 256-neuron fully connected layer + 0.5 Dropout to prevent overfitting
• Sigmoid output layer for binary classification

Data Augmentation Strategy

Data augmentation is emphasized to improve the model's generalization ability, suitable for scenarios with limited medical image data.

Synthetic Dataset Version

Provides synthetic X-ray images for quick testing and debugging. It allows verifying the correctness of the training process without large datasets, facilitating early-stage development iterations.

Real Dataset Version

Supports the Kaggle real chest X-ray dataset (Chest X-ray Pneumonia), which includes training/test/validation sets, classified into normal/pneumonia categories, and provides real medical image data.

The project compares the performance of Adam and SGD optimizers:

• Adam: Fast convergence, higher final accuracy, suitable for scenarios where quick results are needed
• SGD: Slow convergence but stable learning, may have better generalization Developers can choose the training strategy according to their needs.

Provides rich evaluation metrics and visualizations:

• Training curves (accuracy/loss changes)
• Confusion matrix (classification performance)
• Classification report (accuracy, precision, recall, F1)
• Sample prediction visualization
• Temporal performance simulation
• Network architecture diagram Helps understand model behavior and supports academic reports and presentations.

Key Findings

1. Data augmentation significantly reduces overfitting in medical image classification
2. Adam outperforms SGD in this dataset
3. CNN can effectively learn patterns of lung opacity regions related to pneumonia
4. Synthetic datasets shorten the development cycle

Summary

The project has a complete structure and clear documentation, demonstrating the application of CNN in medical image classification. Through dual datasets, optimizer comparison, and visualization outputs, it provides comprehensive references for learners, suitable for beginners in medical AI or computer vision developers to study.

The author proposes the following improvement directions:

• Introduce Grad-CAM interpretability technology
• Use ResNet/EfficientNet pre-trained models to improve performance
• Deploy as a web service using Flask/FastAPI
• Convert to TensorFlow Lite to support mobile devices
• Improve class imbalance handling To further enhance the practical application performance of the model.