Brain Extraction Evaluation for Tumor Segmentation with Deep Learning

An evaluation of different brian extraction tools and methods for brain tumor segmentation using deep learning. This repository accompanies the paper:

B. M. Pacheco, G. de Souza e Cassia, and D. Silva, “Towards fully automated deep-learning-based brain tumor segmentation: Is brain extraction still necessary?,” Biomedical Signal Processing and Control, vol. 82, p. 104514, 2023, doi: https://doi.org/10.1016/j.bspc.2022.104514.

Requirements

Our preprocessing pipeline is implemented using FSL, CaPTk and HD-BET, as detailed in the paper. You will also need nnU-Net and our library of useful functions for brain tumor segmentation (brats), which are included as submodules in this repo. I recommend that you install both in development mode (python setup.py develop) from the submodules, even though everything should work with the nnU-Net package installed from the main source (untested).

Data

We use data from BraTS 2020, which can be acquired in many different ways, and from DICOM-Glioma-SEG, BraTS-TCGA-LGG and BraTS-TCGA-GBM. Note, however, that the BraTS 2018 images of these last two sources are available only under request.

You should unzip BraTS2020 data into data/raw/MICCAI_BraTS2020_TrainingData, TCGA images in data/raw/TCGA, DICOM-Glioma-SEG in data/raw/DICOM-Glioma-SEG and DICOM_Glioma_SEG_Metadata in data/raw/DICOM_Glioma_SEG_Metadata. You must also download the non-skull-stripped T1-weighted image of the SRI24 atlas and store it as data/raw/SRI24_T1.nii.

The preprocessing expects the TCGA data to be organized in a slightly different way from what is provided directly in TCIA. First, the TCGA folder inside data/raw must have two subdirectories, LGG and GBM, containing the respective data for each study. Inside these, the structure is similar to what we get from TCIA, except all whitespaces are converted to underscores. Unfortunately, this is necessary for some of the CLIs of the preprocessores we use. This whitespace-to-underscores conversion is also required for DICOM-Glioma-SEG data. So, for example, the raw dicom files for the T1Gd modality of TCGA-02-0003 patient will be stored in the data/raw/TCGA/GBM/TCGA-02-0003/13.000000-AX_T1_POST-19694/. directory.

BE Methods

As the BE methods require additional instalations, the current implementation applies only FSL's BET and HD-BET, as these installations are already required for the preprocessing pipeline. However, this implementation suits any other BE method. See the template for such in make_test_data.py, the apply_XXX function.

Execution order

With the required packages (environment.yml) and the two submodules installed, run:

1. make_dotenv.py to create a .env file with the environment variables required by nnU-Net package
2. make_nnunet_brats_task.py and make_nnunet_nobe_task.py to preprocess the training data
3. make_test_data.py to preprocess the test data
4. nnunet_pp_tasks.py and nnunet_train_models.py to train the models
5. nnunet_predict.py to use the trained models for prediction
6. compute_results.py to generate a scores.csv file with Dice and HD95 of each image for all models and BE methods