🩺 Pneumonia Detection using Deep Learning (CNN + Transfer Learning + Grad-CAM)

This project builds and evaluates deep learning models to classify chest X-ray images as PNEUMONIA or NORMAL. It explores multiple approaches including shallow CNNs, deep CNNs, transfer learning with MobileNet, and model interpretability using Grad-CAM.

---

🚀 Overview

Pneumonia is a serious lung infection that can be detected using chest X-rays. In this project, we:

• Build CNN models from scratch
• Compare shallow vs deep architectures
• Fine-tune a pretrained model (MobileNet)
• Evaluate performance using ROC curves
• Interpret model predictions using Grad-CAM

---

📊 Dataset

• Dataset: Chest X-ray images
• Classes:
  • NORMAL
  • PNEUMONIA

🧱 Pipeline

1. Data Preprocessing

• Resize images → 112 x 112
• Convert to tensors
• Normalize using ImageNet statistics
• Split training data:
  • 80% train
  • 20% validation

---

🧠 Models Implemented

🔹 1. Shallow CNN (Baseline)

Architecture:

• Conv → ReLU → MaxPool
• Conv → ReLU → MaxPool
• Flatten → Fully Connected

📌 Key Insight:

• Fast to train
• Limited feature extraction capability

---

🔹 2. Deep CNN

Architecture:

• 3 Convolutional layers
• Increasing channels (16 → 32 → 64)
• Dropout for regularization
• Fully connected layers

📌 Key Insight:

• Better feature learning
• Improved generalization vs shallow model

---

🔹 3. Transfer Learning (MobileNet)

• Pretrained on ImageNet
• Replaced final classifier layer
• Fine-tuned on chest X-ray dataset

📌 Key Insight:

• Best performance among all models
• Faster convergence
• Leverages learned features

---

⚙️ Training Details

• Loss Function: CrossEntropyLoss
• Optimizer: Adam
• Learning Rate: 0.0001
• Batch Size: 64
• Epochs: ~20
• Model checkpoint saved every epoch

---

📈 Evaluation Metrics

• ROC Curve
• AUC (Area Under Curve)
• Training vs Validation Loss
• Model generalization analysis

---

📊 Results Summary

Model	Performance
Shallow CNN	Baseline performance
Deep CNN	Improved accuracy
MobileNet	Best performance

📌 Best model: MobileNet (Transfer Learning)

---

📉 Learning Curves

• Training loss decreases steadily
• Validation loss stabilizes early
• Optimal model found around epoch 5–7
• Later epochs show signs of overfitting

---

🔍 Model Interpretability (Grad-CAM)

Grad-CAM is used to visualize where the model is looking when making predictions.

Key Insights:

• Highlights lung regions relevant for pneumonia
• Helps validate model decisions
• Improves trust in predictions

---

🧪 Experimentation

• Compared shallow vs deep architectures
• Evaluated effect of model depth
• Tested transfer learning vs training from scratch
• Analyzed overfitting behavior
• Visualized predictions using Grad-CAM

---

⚡ Challenges

• Overfitting in shallow models
• Limited dataset size
• Class imbalance
• Model interpretability

---

🔮 Future Improvements

• Use larger medical datasets
• Apply data augmentation
• Use advanced architectures (ResNet, EfficientNet)
• Deploy as a web app (Streamlit / FastAPI)
• Add explainability dashboards

---

💻 Tech Stack

• Python
• PyTorch
• Torchvision
• NumPy
• Matplotlib
• Scikit-learn
• pytorch-grad-cam

---

📌 References

• PyTorch Documentation: https://pytorch.org
• Torchvision Models: https://pytorch.org/vision/stable/models.html
• Grad-CAM Paper: https://arxiv.org/abs/1610.02391

---