🧠✨ Synthetic Brain MRI Generation & Classification with AI

📄 Abstract

This project leverages Artificial Intelligence to address ethical and privacy concerns in healthcare by generating synthetic brain cancer images. The aim is to create realistic synthetic datasets that mimic real patient data, enabling effective training of ML models without compromising privacy.

---

🧠 Introduction

Training machine learning models on sensitive healthcare data raises serious ethical and legal issues.
To tackle this, we explored the use of Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN) to generate synthetic brain MRI scans.
We then evaluated the performance of classification models—notably Convolutional Neural Networks (CNN) and VAEs—on real, synthetic, and hybrid datasets.

---

🗂️ Dataset

We used the Br35H dataset, which includes:

• 🧪 3000 MRI scans

  • 🧠 1500 tumor-positive
  • ✅ 1500 tumor-negative

• 📏 Varied image sizes & scanning techniques

• ⚠️ Non-IID structure, posing challenges for consistent model training

🖼️ Example Images

• (a) 'Flair' Brain MRI — Size: 587×630

• (b) 'T2' Brain MRI — Size: 197×256

---

🧪 Methodology

🧼 Pre-processing

• Resized and normalized all images to [0, 1]

• Maintained aspect ratios

• Unified backgrounds and dimensions

🧬 Data Generation

• GAN (Generative Adversarial Network)

  • Generator vs Discriminator in a min-max game

• VAE (Variational Autoencoder)

  • Encoder-decoder architecture mapping to latent space

🔁 Trained separately for tumor-positive and tumor-negative images to ensure balance

🧠 Classification

🔍 Models

• CNN Architecture: Feature extraction + classification layers

• VAE for Classification: Used reconstruction error as a signal for anomaly (tumor) detection

🛠️ Techniques

• 📏 Global Image Absolute Error Magnitude (GIAEM)

• 🌐 DBSCAN: Density-based spatial clustering

• 🎯 Singular-example KMeans

• 📊 Global KMeans: Clustered reconstruction errors across dataset

---

📊 Experimental Results

🎨 Generative Task

Model	Result
GAN	❌ Struggled with realistic tumor generation, high noise
VAE	✅ More realistic images, but difficulties in tumor regions

🤖 Classification Task

Model	Performance
CNN	🟢 High accuracy on real data, poor generalization to synthetic
VAE	🟡 Varying results; Global KMeans achieved best balance

---

📈 Performance Metrics

Model	Train Data	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
CNN	Real	96.67	95.42	97.99	96.69
	Synthetic	58.16	57.53	56.94	57.24
	Mixed	49.97	49.98	99.93	66.64
VAE (GIAEM)	Real	60.44	60.95	67.11	56.76
	Synthetic	80.08	68.26	66.22	67.09
	Mixed	81.24	69.99	64.11	65.95
VAE (Global KMeans)	Real	82.93	73.55	67.58	69.66
	Synthetic	79.39	68.26	69.44	68.80
	Mixed	78.80	39.88	49.27	44.08

---

🧾 Conclusions & Future Work

• ⚠️ The Br35H dataset posed difficulties due to its heterogeneous nature

• 🔍 VAE proved to be the more effective generative model

• 🧠 CNN excelled on real data but lacked generalization

• 🔬 Future work should:

  • Explore more consistent datasets
  • Enhance GAN performance
  • Expand unsupervised techniques like DBSCAN and KMeans

---

📚 References

• CovidGAN: Data Augmentation Using Auxiliary Classifier GAN for Improved Covid-19 Detection – Abdul Waheed et al.

• Synthetic Medical Images Using F&BGAN for Improved Lung Nodules Classification – Defang Zhao et al.

• Combining Noise-to-Image and Image-to-Image GANs: Brain MR Image Augmentation for Tumor Detection

• Infinite Brain MR Images: PGGAN-based Data Augmentation for Tumor Detection

• Br35H: Brain Tumor Detection 2020

• DBSCAN

• KMeans

---

👨‍💻 Team

• 🧬 Generation Team: Francesco D'Aprile, Sara Lazzaroni

• 🔎 Classification Team: Anthony Di Pietro, Tommaso Mattei

---

Sources

At the following Google Drive link, you can find the dataset (both real and synthetic) and the weights for the classification and generation models. https://drive.google.com/drive/folders/1JJT4MP_5GSH_CU1Blvm6GpfhzZvHZFlb

📎 More Info

🔗 Check out the full paper in doc/ for more details!

Built With

• python
• pytorch
• scikit-learn