Privacy-Preserving Brain Tumor Analysis with Federated Learning & Explainable AI

A research project exploring privacy-preserving medical image analysis using Federated Learning, Byzantine-robust aggregation, and Explainable AI (Grad-CAM) for trustworthy brain tumor detection from MRI scans.

---

📖 Overview

Medical imaging datasets contain highly sensitive patient information, making centralized model training difficult due to privacy regulations. Federated Learning (FL) enables multiple healthcare institutions to collaboratively train a shared model while keeping patient data local.

This repository presents a two-stage project:

1. Brain Tumor Detection with CNN & Grad-CAM – a centralized deep learning model for MRI classification with visual explainability.
2. Privacy-Preserving Brain Tumor Analysis – an extension of the baseline model into a federated learning framework that simulates distributed hospital training and evaluates robust aggregation methods against malicious clients.

Together, these projects demonstrate how explainable and privacy-preserving AI can be combined for trustworthy medical image analysis.

---

✨ Features

• 🧠 Brain tumor detection from MRI images
• 🏥 Federated Learning across simulated hospital clients
• 🔒 Privacy-preserving distributed training
• 🛡️ Byzantine-robust aggregation using Trimmed Mean
• 📊 FedAvg vs. Trimmed Mean comparison
• 🔥 Grad-CAM explainability for MRI interpretation
• 📈 Performance evaluation under centralized and federated settings
• 🤖 Built with TensorFlow/Keras

---

📂 Repository Structure

Privacy-Preserving-Brain-Tumor-Analysis/
│
├── Brain_Tumor_Detection.ipynb
│   ├── CNN Model
│   ├── MRI Classification
│   └── Grad-CAM Visualization
│
├── BrainTumor_Federated_Byzantine.ipynb
│   ├── Federated Learning Simulation
│   ├── FedAvg Aggregation
│   ├── Trimmed Mean Aggregation
│   ├── Byzantine Client Simulation
│   └── Performance Evaluation
│
├── images/
│   ├── gradcam_example.png
│   ├── federated_setup.png
│   └── results.png
│
├── requirements.txt
├── LICENSE
└── README.md

---

🚀 Project Workflow

Brain MRI Dataset
        │
        ▼
Brain_Tumor_Detection.ipynb
        │
        ├── CNN Training
        ├── Model Evaluation
        ├── MRI Classification
        └── Grad-CAM Explainability
        │
        ▼
BrainTumor_Federated_Byzantine.ipynb
        │
        ├── Simulated Hospital Clients
        ├── Local Training
        ├── FedAvg Aggregation
        ├── Trimmed Mean Aggregation
        ├── Byzantine Attack Simulation
        └── Federated Model Evaluation

---

🧠 Dataset

Source

• Kaggle – Brain MRI Images for Brain Tumor Detection

Property	Value
Total Images	253
Tumor Images	155
Healthy Images	98
Classes	2
Image Type	Brain MRI

---

📚 Project 1 — Brain Tumor Detection

This notebook implements a CNN-based classifier for detecting brain tumors from MRI scans.

Model Architecture

Input (128×128×3)
        │
Conv2D (32)
        │
MaxPooling
        │
Conv2D (64)
        │
MaxPooling
        │
Flatten
        │
Dense (128)
        │
Dropout
        │
Dense (2)

Training Configuration

Parameter	Value
Optimizer	Adam
Loss	Categorical Crossentropy
Epochs	10
Batch Size	32
Image Size	128×128
Train/Test Split	80/20

Data Augmentation

• Random Rotation
• Zoom
• Width Shift
• Height Shift
• Horizontal Flip

---

🔥 Explainable AI with Grad-CAM

To improve model transparency, Grad-CAM is used to visualize which regions of an MRI most influenced the prediction.

The Grad-CAM pipeline:

1. Extract feature maps from the final convolutional layer.
2. Compute gradients of the predicted class.
3. Weight feature maps using pooled gradients.
4. Generate a heatmap.
5. Overlay the heatmap on the original MRI.

This helps clinicians verify whether predictions are based on medically meaningful regions.

Example implementation:

with tf.GradientTape() as tape:
    conv_outputs, predictions = grad_model(img_array)
    loss = predictions[:, tf.argmax(predictions[0])]

grads = tape.gradient(loss, conv_outputs)
pooled_grads = tf.reduce_mean(grads, axis=(0,1,2))
heatmap = conv_outputs[0] @ pooled_grads[..., tf.newaxis]

---

🏥 Project 2 — Privacy-Preserving Federated Learning

The second notebook extends the centralized model into a federated learning framework.

Instead of pooling MRI scans into one dataset, multiple simulated hospitals collaboratively train a shared model while keeping patient data local.

Federated Learning Setup

Hospital A
      │
Hospital B
      │
Hospital C
      │
Hospital D
      │
      ▼
 Federated Server
      │
      ▼
 Global Model

Each hospital:

• Trains the model locally
• Never shares patient MRI images
• Sends only model updates to the central server

---

🛡️ Robust Aggregation

Two aggregation strategies are evaluated.

FedAvg

Standard Federated Averaging.

Advantages

• Simple
• Efficient
• Widely used

Limitations

• Sensitive to malicious or poisoned client updates

---

Trimmed Mean

A Byzantine-robust aggregation algorithm that removes extreme client updates before averaging.

Advantages

• Resistant to malicious updates
• More stable under adversarial settings
• Better suited for safety-critical applications

---

📊 Results

Centralized CNN

Metric	Value
Test Accuracy	~70.6%

The relatively small dataset limits performance, providing opportunities for improvement through larger datasets or transfer learning.

---

Federated Learning

The federated learning experiments compare:

• FedAvg
• Trimmed Mean

Evaluation focuses on:

• Classification performance
• Privacy-preserving distributed learning
• Robustness against malicious clients
• Explainability using Grad-CAM

---

💻 Technologies Used

• Python
• TensorFlow / Keras
• NumPy
• OpenCV
• Scikit-learn
• Matplotlib

---

🚀 Future Work

Potential extensions include:

• Differential Privacy
• Secure Aggregation
• Vision Transformers (ViT)
• Flower or FedML frameworks
• Larger medical imaging datasets (BraTS)
• Multi-class brain tumor classification
• Self-supervised medical vision models

---

📖 Acknowledgements

• Dataset: Brain MRI Images for Brain Tumor Detection (Kaggle)
• Grad-CAM: Selvaraju et al., Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization, ICCV 2017.

---

📄 License

This project is licensed under the MIT License.