joint_registration_tumor_segmentation

This repository is the official repository of the article Deep Learning-Based Concurrent Brain Registration and Tumor Segmentation

The article is available here : https://www.frontiersin.org/articles/10.3389/fncom.2020.00017/abstract

To train a model use the main.py function. The most important arguments are :

• --segmentation-only to train only on the segmentation task
• --registration-only to train only on the registration task
• --with-loss-trick to add the seg^0 (Equation 4 of the article)
• --only-t1 to train only with t1 modality (by default 4 modality are excepted)
• --source-target--merge-operation ['subtraction', 'addition', 'concatenation'] to choose the mergin operation used

To train the proposed method, the commands line is :

python -m joint_registration_tumor_segmentation.main --only-t1 --session-name seg_reg_SubMerge_8channels_1.0ratio_0.002lr --epochs 180 --batch-size 2 --lr 2e-3 --nb-gpu 1 --only-brats --source-target-merge-operation subtraction --n-channels-first-layer 8 --ratio-weights-registration-over-segmentation 1.0 -deform-regularisation 1e-10

To do the inference, use the inference.py function. The most important arguments are :

• --get-segmentation to save the predicted segmentation
• --get-registration to save the predicted registration
• --model-abspath to give the absolue path of one model to do the inference with
• --models-folder to give the absolue path of one folder containing different models to do the inference with

To do the inference, the command line is :

python -m joint_registration_tumor_segmentation.inference  --get-registration --data-folder-path path_to_data --model-abspath path_to_model --output-folder path_to_output

Data

In this article we use two datasets : OASIS 3 and BraTS 2018. BraTS 2018 was used for training and testing both segmentation and registration. OASIS 3 was used only for testing. More details are given in the article.

We don't provide the datasets in this repo. People can find the datasets on the following link : https://www.oasis-brains.org/ and https://www.med.upenn.edu/sbia/brats2018/data.html

To run the code, you need to download both datasets. By default, the data is supposed to be in the folder joint_registration_tumor_segmentation\data\oasis\ and joint_registration_tumor_segmentation\data\BRATS\