Brain tumor segmentation on MRI image

We will build segmentation networks to detect tumor of a brain on the MRI image.

Network architectures which are used is Deeplab v3+ with resnet 101 backbone and Unet.

Getting Started

Before start training, let prepare environment and data

Prerequisites

Installing env

We recommend you to use Virtualenv to manage your libraries

• Python 3.5 or Python 3.6
• File requirements.txt include all libraries

Install libraries:

pip install -r requirements.txt

Prepare data

Download data from Kaggle

Extract it to a folder.

Create train, validation, test set with below command

python data_preprocess/split_data.py --root_dir xxx --output_link yyy

• root_dir: directory include all image original and label after extract data from zip file
• output_link: link to directory to save file train.txt, vali.txt, test.txt

Output folder will have structures like below:

Files train.txt, vali.txt, test.txt include lines. Each lines compose sub-links to file image origin and image label with tab separate

Example

TCGA_CS_4941_19960909/TCGA_CS_4941_19960909_1.tif   TCGA_CS_4941_19960909/TCGA_CS_4941_19960909_1_mask.tif
TCGA_CS_4941_19960909/TCGA_CS_4941_19960909_2.tif   TCGA_CS_4941_19960909/TCGA_CS_4941_19960909_2.tif

Running the training and inference

To training run python train.py --

python train.py --root_folder /root/data/kaggle_3m \
                --train_data /root/data/train.txt \
                --validation_data /root/data/vali.txt \
                --trainer_save_dir /root/data/save_dir

• root_folder: folder include image origin and image label
• train_data: file include link to train data
• validation_data: file include link to validation data
• trainer_save_dir: Folder where results(log, checkpoint, tensorboard) will be saved.

Results:

I got 0.815 IoU with Deeplab v3+ resnet 101 and 0.796 with Unet.

Some inference image segment by deeplab:

Replace data with other binary dataset(lung segmentation):

Contributing

Authors

License

This project is licensed under the MIT License

Acknowledgments