Brain Tumor Detection — GAN + ResNet50

A deep learning pipeline for binary brain tumor classification from MRI scans, combining Generative Adversarial Networks (GAN) for data augmentation with a ResNet50 transfer learning classifier.

This project is an implementation based on the research published in:

MV Sujan Kumar, Ganesh Khekare, Shashi Kant Gupta, and Sharnil Pandya. Harnessing generative AI for enhanced brain tumor detection in clinical trials. In Generative AI Unleashed, Chapter 6, pp. 83–101, IET, 2025. https://doi.org/10.1049/PBPC076E_ch6

---

Overview

Brain tumor detection from MRI scans is a critical clinical task where dataset size is a constant bottleneck. This project addresses class imbalance and data scarcity using a GAN-based augmentation strategy, feeding synthetic MRI images alongside real ones into a frozen ResNet50 backbone for binary classification (Tumor / No Tumor).

---

Architecture

Brain MRI Dataset (253 images)
         │
         ▼
  Data Augmentation
  (flips → 1,000 images/class)
         │
    ┌────┴────┐
    │         │
    ▼         ▼
GAN Training  Real Images
(200 epochs)  (normalised)
    │         │
    ▼         │
Generator     │
Conv2DTranspose ×3
tanh output   │
    │         │
    ▼         ▼
Synthetic   Train Split
MRI Scans   (70%)
    │         │
    └────┬────┘
         │
         ▼
  Combined Training Set
  (Real + GAN-augmented)
         │
         ▼
  ResNet50 (frozen, ImageNet)
         │
    Flatten
         │
    Dense(1024) + Dropout(0.4)
         │
    Dense(1, sigmoid)
         │
         ▼
  Tumor / No Tumor

GAN Architecture

Component	Details
Input	Latent vector (dim=100)
Generator	Dense → Reshape → Conv2DTranspose ×3 → tanh
Discriminator	Conv2D ×2 + LeakyReLU(0.2) + Dropout → Dense(1, sigmoid)
Output size	32×32 grayscale images
Training	200 epochs, label smoothing 0.9, Adam (lr=0.0002)

Classifier Architecture

Component	Details
Backbone	ResNet50 (frozen, ImageNet weights)
Input size	256×256×3
Head	Flatten → Dense(1024, ReLU) → Dropout(0.4) → Dense(1, sigmoid)
Training	10 epochs, Adam, binary cross-entropy
Train/Test split	70% / 30%

---

Results

Metric	Value
Accuracy	98.87%
Precision	93.74%
Recall	92.14%
F1 Score	92.93%
AUC-ROC	0.96

Confusion Matrix

ROC Curve

---

Dataset

Kaggle — Brain MRI Images for Brain Tumor Detection

Split	Normal	Tumor
Raw	98	155
After augmentation	1,000	1,000
Train	—	708 total
Test	—	304 total

Dataset not included in this repository. Download and place images in data/yes/ and data/no/.

---

Project Structure

brain-tumor-gan-resnet/
├── app/
│   └── app.py              # Gradio inference app
├── assets/
│   ├── confusion_matrix.png
│   └── roc_curve.png
├── data/
│   ├── yes/                # Tumor MRI images (not included)
│   ├── no/                 # Normal MRI images (not included)
│   └── README.md
├── notebooks/
│   └── brain_tumor_detection.ipynb
├── outputs/
│   ├── models/             # Saved .keras models (not included)
│   └── plots/
├── src/
│   ├── config.py
│   ├── data_loader.py
│   ├── gan.py
│   ├── classifier.py
│   ├── train.py
│   └── evaluate.py
├── requirements.txt
├── LICENSE
└── README.md

Source code is intentionally omitted from this repository. The architecture, methodology, and results are documented here for portfolio and research reference.

---

Citation

@inbook{doi:10.1049/PBPC076E_ch6,
  author    = {MV Sujan Kumar and Ganesh Khekare and Shashi Kant Gupta and Sharnil Pandya},
  title     = {Harnessing generative AI for enhanced brain tumor detection in clinical trials},
  booktitle = {Generative AI Unleashed},
  chapter   = {Chapter 6},
  pages     = {83--101},
  doi       = {10.1049/PBPC076E_ch6},
  url       = {https://digital-library.theiet.org/doi/abs/10.1049/PBPC076E_ch6},
  year      = {2025},
  publisher = {Institution of Engineering and Technology}
}

---

License

This repository is licensed under the MIT License.

The implementation is based on published research. All academic credit belongs to the original authors cited above.