Brain Tumor Segmentation

This repository contains a deep learning model for brain tumor segmentation from MRI images using a U-Net architecture. The model is trained on the LGG MRI Segmentation dataset from Kaggle.

The project implements a U-Net model to segment brain tumors from MRI scans. The model takes MRI images as input and outputs a binary mask indicating the presence and location of tumors.

Dataset

The model uses the LGG MRI Segmentation dataset from Kaggle.

• Contains 3,929 MRI images in TIFF format with corresponding masks
• Masks are binary (0 for no tumor, 1 for tumor present)

Data Preprocessing

The data preprocessing pipeline includes:

1. Downloading and extracting the dataset from Kaggle
2. Organizing the data into a pandas DataFrame with columns:
   • patient_id
   • image_path
   • mask_path
   • mask (binary indicator)
3. Splitting the data into training, validation, and test sets.
4. Creating a custom data generator (BrainTumorGenerator) that:
   • Loads images and masks
   • Resizes them to 256x256 pixels

Model Architecture

The model uses a U-Net architecture with the following components:

Encoder (Downsampling Path):

• 3 levels of convolutional blocks (Conv2D + Conv2D + MaxPooling2D)
• Number of filters increases from 64 to 256

Bottleneck:

• Two convolutional layers with 512 filters

Decoder (Upsampling Path):

• 3 levels of upsampling blocks (UpSampling2D + concatenation + Conv2D + Conv2D)
• Skip connections from corresponding encoder levels
• Number of filters decreases from 256 to 64

Output:

• Final 1x1 convolution with sigmoid activation for binary segmentation

Training

The model was trained with:

• Optimizer: Adam
• Loss function: Binary crossentropy
• Batch size: 16
• Image size: 256x256
• Callbacks:
  • Early stopping (patience=3)
  • Model checkpointing (saves best model)

Model Performance

The model achieved the following performance:

• Validation Accuracy: 99.41%
• Test Accuracy: 99.10%

The model demonstrates excellent performance in segmenting brain tumors from MRI images, with high accuracy on both validation and test sets.

Requirements

To run this project, you'll need:

• numpy
• pandas
• opencv
• pillow
• matplotlib
• scikit-learn
• tensorflow
• kaggle (for dataset)