Chest X-ray Classification (COVID-19, Pneumonia, Lung Opacity, Normal)

This project focuses on classifying chest X-ray images into four classes: COVID-19, Viral Pneumonia, Lung Opacity, and Normal. It explores deep learning techniques like CNN and Transfer Learning under the challenges of medical image classification and class imbalance.

---

🗂 Dataset

• Source: Kaggle chest X-ray dataset
• ~19,000 X-ray images (unbalanced across classes)
• Includes segmentation masks for lungs (used for region-focused training)

Preprocessing:

• Applied lung segmentation masks to focus on regions of interest
• Data augmentations: mild random rotations and flips

---

🧠 Models

1. CNN Baseline

• Custom convolutional network trained from scratch
• Used as a performance baseline

2. ResNet18 (Transfer Learning)

• Pretrained on ImageNet
• Fully fine-tuned
• Compared with and without class imbalance strategies

---

⚖️ Class Imbalance Strategies

1. None (Baseline)
2. Weighted Loss: Class weights inversely proportional to frequency
3. Oversampling: WeightedRandomSampler to balance batches

---

📊 Evaluation Metrics

• Per-class: Precision, Recall, F1-Score
• Macro averages across all classes
• Accuracy (for reference)
• Confusion matrix
• ROC and Precision-Recall curves (per class)

---

✅ Results (Macro F1-Score)

Model	Weighted Loss	Oversampling	Macro F1
CNN Baseline	No	No	0.80
ResNet18	No	No	0.89
ResNet18	Yes	No	0.86
ResNet18	No	Yes	0.90

Note: All metrics and plots you can find in notebooks/results.ipynb

---

🏁 Conclusion

• Custom CNN provides a reasonable starting point.
• Transfer learning with ResNet18 outperforms it.
• Oversampling gave the best results in this setup.
• Class Imbalance Strategies was not that effective (minor class had already good scoring)

---

🛠️ Requirements

python >= 3.8
torch >= 1.10
torchvision
scikit-learn
matplotlib
numpy
pandas

---

📁 Structure

├── models/                # CNN baseline architecture + save/load model functions
├── notebooks/             # Training and results with visualizations
├── outputs/               # Saved curves and confusion matrix as .png files
├── preprocessing/         # Creating DataLoaders, Data Augmentation, Class Imbalance Strategies functions
├── results/               # Evaluation + Visualizations functions
├── training/              # Training loops
├── utils/                 # Other helper functions
├── weights/               # Saved model's weights
├── main.py                # Pipeline to train and evaluate model