Brain Tumor Segmentation and Classification

A multi-task deep learning model for simultaneous brain tumor segmentation and classification using the BRISC2025 dataset. This implementation uses a lightweight U-Net architecture optimized for Google Colab training.

Overview

This project implements a dual-output neural network that performs:

• Segmentation: Pixel-level tumor boundary detection
• Classification: Tumor type identification across 4 categories

The model achieves strong performance with optimized training parameters designed for efficient Colab execution.

Dataset

BRISC 2025 (Brain Tumor MRI Dataset for Segmentation and Classification)

A high-quality, expert-annotated MRI dataset addressing common limitations in existing datasets like BraTS and Figshare. The dataset includes class imbalance corrections, comprehensive tumor coverage, and consistent radiologist-verified annotations.

Dataset Specifications:

• 6,000 T1-weighted MRI images with corresponding segmentation masks
• Four tumor classes: Glioma, Meningioma, Pituitary, and No Tumor
• Multiple anatomical views: axial, coronal, sagittal
• Expert-curated annotations by radiologists and physicians
• ArXiv preprint: https://arxiv.org/abs/2506.14318

Tumor Classes:

• gl: Glioma
• me: Meningioma
• pi: Pituitary
• nt: No Tumor

Architecture

Lightweight Multi-Task U-Net

The model features a shared encoder with two specialized decoder branches:

Encoder:

• Three encoder blocks with progressively increasing filters (32, 64, 128)
• Max pooling for downsampling
• Batch normalization and ReLU activation

Segmentation Branch:

• Three decoder blocks with skip connections
• Transposed convolutions for upsampling
• Binary mask output with sigmoid activation

Classification Branch:

• Global average pooling on bottleneck features
• Two dense layers (128, 64 units) with dropout
• Softmax output for 4-class prediction

Model Optimizations:

• Reduced filter counts for faster training
• Smaller input resolution (128x128)
• Lightweight classification head
• Efficient skip connections

Training Configuration

Hyperparameters:

Image size: 128x128
Batch size: 16
Epochs: 50
Learning rate: 5e-4
Validation split: 15%

Loss Functions:

• Segmentation: Dice loss
• Classification: Categorical cross-entropy

Data Augmentation:

• Random horizontal flipping
• Random brightness adjustment

Results

The model demonstrates excellent performance across both tasks:

Segmentation Metrics:

Metric	Score
Dice Coefficient	0.7814
IoU	0.6975
Pixel Accuracy	0.9944
Sensitivity	0.7908
Specificity	0.9978

These metrics indicate strong tumor boundary detection with high precision and minimal false positives.

Setup and Installation

Requirements:

tensorflow
numpy
opencv-python
scikit-learn

Google Colab Setup:

1. Mount Google Drive:

from google.colab import drive
drive.mount('/content/drive')

2. Ensure dataset is organized in the following structure:

Brain Tumor Segmentation and Classification/
└── training_SupervisedLearning/
    └── brisc2025/
        ├── segmentation_task/
        │   └── train/
        │       ├── images/
        │       └── masks/
        └── classification_task/

3. Run the training script in a Colab notebook cell

Usage

Training:

Simply execute the main script:

python train.py

The training process will:

• Load and split the dataset
• Display class distribution statistics
• Build the multi-task model
• Train with automatic checkpointing
• Save the best model and training history

Output Files:

All outputs are saved to the checkpoints directory:

• Best model checkpoint based on validation loss
• Final trained model
• Training history in JSON format

Training Monitoring:

The custom callback displays epoch-by-epoch metrics:

• Segmentation loss and Dice coefficient
• Classification loss and accuracy
• Validation metrics for both tasks

Model Performance

Training achieves convergence within 50 epochs with early stopping monitoring. The multi-task learning approach enables the model to leverage shared representations, improving both segmentation and classification performance.

Key Performance Indicators:

• High Dice coefficient indicates accurate tumor boundary detection
• Near-perfect specificity minimizes false tumor predictions
• Balanced sensitivity ensures most tumors are correctly identified
• Strong pixel accuracy reflects overall segmentation quality

File Naming Convention

Dataset files follow the BRISC2025 format:

brisc2025_train_00001_gl_ax_t1.jpg

Components:

• Dataset identifier: brisc2025
• Split: train/test
• Index: 00001
• Tumor code: gl/me/pi/nt
• View: ax/co/sa (axial/coronal/sagittal)
• Sequence: t1 (T1-weighted)

Optimization Features

This implementation includes several optimizations for Colab training:

• 4x faster processing with reduced image resolution
• 2x larger batch size for better GPU utilization
• 50% reduction in model parameters
• Simplified architecture with 3 encoder layers
• Efficient data generators with on-the-fly loading
• Stratified train-validation split

Next Steps

After training completion:

1. Run validation analysis on the saved model
2. Evaluate performance on the test set
3. Visualize predictions on sample images
4. Export model for deployment or inference

License

Please refer to the BRISC2025 dataset documentation for usage terms and conditions.

Citation

If you use this code or the BRISC2025 dataset, please cite:

@article{brisc2025,
  title={BRISC 2025: Brain Tumor MRI Dataset for Segmentation and Classification},
  journal={ArXiv preprint},
  url={https://arxiv.org/abs/2506.14318},
  year={2025}
}

Acknowledgments

Dataset curated and annotated by expert radiologists and physicians. Model architecture based on the U-Net framework adapted for multi-task learning.