Lung Disease Classification - Experimentation Repository

A comprehensive experimentation repository exploring various deep learning and traditional machine learning approaches for automated lung disease classification from chest X-ray images.

Overview

This repository contains extensive experiments comparing traditional machine learning methods with state-of-the-art deep learning architectures for classifying chest X-ray images into three categories:

• Normal - Healthy lung conditions
• Lung Opacity - Various degrees of lung abnormalities
• Viral Pneumonia - Viral pneumonia infections

Live Demo

See the best model in action. The Custom ANN (91.2% accuracy) is deployed as an interactive web application.

Try it: https://lung-disease-classification.streamlit.app/

Web App Repository: https://github.com/laksh-ya/Lung-Disease-Classification/

Dataset

Source: Kaggle Lung Disease Dataset

Composition:

• Total Images: 3,475 chest X-ray images
• Normal: 1,250 images
• Lung Opacity: 1,125 images
• Viral Pneumonia: 1,100 images

Experimental Pipeline

Traditional Machine Learning

• Feature Extraction: SIFT, HOG, LBP
• Dimensionality Reduction: PCA
• Classifier: Logistic Regression
• Result: 89.78% accuracy

Deep Learning Models Tested

Model	Accuracy (%)
ResNet50 (Version 1)	88.27
ResNet50 (Version 2)	89.17
EfficientNetB0	88.87
DenseNet	85.11
Custom CNN	88.72
Custom ANN	88.72
Custom ANN (Revised)	91.2

Optimization Techniques Applied

• SMOTE - Synthetic Minority Over-sampling for class balancing
• Hyperparameter Tuning - Learning rate, batch size, epochs optimization
• Transfer Learning - Fine-tuning pre-trained models
• Explainability - LIME and SHAP for model interpretability

Best Model Performance

Custom ANN (Revised) achieved the highest accuracy:

Class	Precision	Recall	F1-Score	Support
Normal	0.92	0.94	0.93	418
Abnormal	0.90	0.87	0.88	247
Overall Accuracy			0.91	665

• Macro Average: Precision 0.91, Recall 0.90, F1-Score 0.91
• Weighted Average: Precision 0.91, Recall 0.91, F1-Score 0.91

Repository Contents

This repository contains multiple Jupyter notebooks exploring:

• Traditional ML approaches (SIFT + HOG + LBP)
• Deep learning architectures (ResNet, EfficientNet, DenseNet)
• Custom CNN and ANN implementations
• SMOTE data balancing experiments
• Model explainability visualizations (LIME, SHAP)
• Comparative analysis notebooks

Technologies Used

• Deep Learning: TensorFlow, Keras
• Computer Vision: OpenCV, scikit-image
• Feature Extraction: SIFT, HOG, LBP
• Data Processing: NumPy, Pandas, scikit-learn
• Balancing: SMOTE (imbalanced-learn)
• Explainability: LIME, SHAP
• Visualization: Matplotlib, Seaborn

Key Findings

1. Deep learning outperforms traditional ML - Custom ANN (91.2%) vs Traditional ML (89.78%)
2. SMOTE improves performance - Balancing reduces bias toward majority class
3. Hyperparameter tuning is crucial - Revised ANN gained 2.5% accuracy improvement
4. Explainability builds trust - LIME/SHAP highlight critical diagnostic regions
5. Transfer learning shows promise - Pre-trained models provide competitive baselines

Related

• Deployment App: Lung Disease Classifier
• Model Weights: HuggingFace

---