🫁 Chest X-Ray Pneumonia Detection

This project is focused on detecting Pneumonia from Chest X-Ray images using different deep learning architectures: CNN, ResNet, Graph Neural Networks (GNN), and a Multimodal Deep Learning approach inspired by recent research.

---

📁 Dataset

I used the Chest X-Ray Pneumonia dataset from Kaggle:
🔗 Chest X-Ray Images (Pneumonia) | Kaggle

• The dataset includes labeled chest X-ray images:
  • Pneumonia
  • Normal
• It is imbalanced, with more pneumonia images than normal, so special care was taken during training (e.g., augmentation).

---

🧠 Models Implemented

1️⃣ CNN (Convolutional Neural Network)

📈 Accuracy: ~92% (Test)

Highlights:

• Implemented a custom CNN from scratch.
• Faced overfitting initially → overcame it with:
  • Dropout
  • L2 Regularization
  • EarlyStopping
• Achieved excellent performance in:
  • Precision
  • Recall
  • F1 Score

✅ CNN proved to be the best performing model in this project.

---

2️⃣ ResNet (Transfer Learning)

📈 Accuracy: ~85% (Test)

Highlights:

• Used ResNet with ImageNet pretrained weights.
• Experimented with:
  • Freezing and unfreezing layers
  • Fine-tuning to adapt to medical domain
• Still has room for improvement via:
  • Advanced data augmentation
  • Exploring learning rate schedules and optimizers

---

3️⃣ GNN (Graph Neural Network)

Model: Graph Attention Network (GAT)

Highlights:

• Used the repo: Chest-X-Ray-Detection
• Converted image data into graph-structured input for GAT
• Aimed to capture relational dependencies between image patches(used SLIC method-Simple Linear Iterative Clustering)
• Interesting experiment showing how graph-based reasoning can be applied in medical imaging

---

4️⃣ Multimodal Approach (Inspired by Research Paper)

🔗 Multimodal CNN - Kaggle Notebook

Highlights:

• Inspired by cross-modal contrastive learning techniques
• Extracted:
  • Scaled WSI (Whole Slide Image)
  • Image Patches
• Applied a contrastive module for multimodal fusion

🖼️ Multimodal Architecture Example:

---

🛠️ Technologies Used

• Python, NumPy, Pandas
• PyTorch, TensorFlow/Keras
• Torch Geometric (for GNN)
• OpenCV, Matplotlib
• Kaggle Kernels for training and evaluation

---

✅ Results Summary

Model	Accuracy	Highlights
CNN	92%	Best model, overcame overfitting
ResNet	85%	Transfer learning with ImageNet weights
GNN (GAT)	82%	Graph-based patch learning
Multimodal	83.5%	Combined global & patch views with contrastive learning

👨‍💻 Author

Rishit Kar
📧 LinkedIn
🧠 AI | 💻 ML | 📊 Research | 🔬 Deep Learning