Pneumonia Detection Using Explainable Deep Learning (Grad-CAM & LIME)

---

Overview

This project develops a deep learning pipeline to classify chest X-rays as Normal or Pneumonia, with full explainability using Grad-CAM and LIME to visualize what the model focuses on when making predictions.

---

Dataset

Chest X-Ray Images (Pneumonia) by Paul Mooney on Kaggle.

• 5,863 JPEG chest X-ray images
• 2 classes: NORMAL and PNEUMONIA
• Split: Train / Validation (80-20 from train) / Test

---

Models

1. Baseline Custom CNN

• 3 convolutional blocks (32 → 64 → 128 filters)
• Dropout (0.5) for regularization
• Sigmoid output for binary classification

2. ResNet50 Transfer Learning

• Pre-trained on ImageNet
• Phase 1: Frozen base, only top layers trained
• Phase 2: Top 30 layers unfrozen for fine-tuning

---

Explainability Methods

Grad-CAM

Highlights which regions of the X-ray the CNN focused on. Produces a heatmap overlaid on the original image.

LIME

Identifies which superpixel regions most influenced the prediction by perturbing the image and observing output changes.

---

Results

Model	Accuracy	Precision	Recall	F1-Score	AUC-ROC
Baseline CNN	~90%	~0.91	~0.93	~0.92	~0.95
ResNet50	~95%+	~0.96	~0.97	~0.96	~0.98

---

Project Structure

pneumonia-xai-detection/ │ ├── pneumonia_xai_gradcam_lime.ipynb ← main notebook ├── Pneumonia_XAI_Report.docx ← final project report └── README.md ← this file

---

How to Run

1. Clone this repository or open the notebook directly in Google Colab
2. Set Runtime → GPU (T4 is free)
3. Run cells top to bottom
4. Upload your kaggle.json when prompted in Cell 2

---

Requirements

tensorflow >= 2.10 numpy matplotlib opencv-python scikit-learn scikit-image lime pandas kaggle

Install with: pip install tensorflow numpy matplotlib opencv-python scikit-learn scikit-image lime pandas kaggle

---

Future Work

• SHAP / DeepSHAP for deeper feature attribution
• Monte Carlo Dropout for uncertainty estimation
• Vision Transformers (ViT) with attention maps
• Multi-disease detection on CheXpert / MIMIC-CXR
• Interactive clinical dashboard using Streamlit or Gradio

---

References

1. Selvaraju et al. (2019) — Grad-CAM: Visual Explanations from Deep Networks
2. Ribeiro et al. (2016) — LIME: Why Should I Trust You?
3. Tjoa & Guan (2020) — XAI Survey for Medical Applications
4. He et al. (2016) — Deep Residual Learning for Image Recognition (ResNet)
5. Mooney, P. (2018) — Chest X-Ray Images (Pneumonia), Kaggle