Automatic Vestibular Schwannoma Segmentation using Deep Learning

This repository provides scripts for the segmentation of vestibular schwannoma on 3D T1 or T2 images, based on the following papers:

• Shapey, J. et al. Segmentation of vestibular schwannoma from MRI, an open annotated dataset and baseline algorithm. Scientific Data, 8(1), p.286. https://doi.org/10.1038/s41597-021-01064-w
• Wang, G. et al. Automatic Segmentation of Vestibular Schwannoma from T2-Weighted MRI by Deep Spatial Attention with Hardness-Weighted Loss, MICCAI, pp 264-272, 2019.
• Shapey, J. et al. An artificial intelligence framework for automatic segmentation and volumetry of vestibular schwannomas from contrast-enhanced T1-weighted and high-resolution T2-weighted MRI. Journal of Neurosurgery, 1(aop), pp.1-9.

TCIA dataset

The dataset described in these publications is publicly available in The Cancer Imaging Archive (TCIA):

Shapey, J., Kujawa, A., Dorent, R., Wang, G., Bisdas, S., Dimitriadis, A., Grishchuck, D., Paddick, I., Kitchen, N., Bradford, R., Saeed, S., Ourselin, S., & Vercauteren, T. (2021). Segmentation of Vestibular Schwannoma from Magnetic Resonance Imaging: An Open Annotated Dataset and Baseline Algorithm [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/TCIA.9YTJ-5Q73

Note, that one subject (VS-SEG-168) had to be removed from the data set because the data was incomplete.

TCIA offers to download the data set under a "Desciptive Folder Name" or a "Classic Folder Name". We recommend choosing the Descriptive Folder Name and using the instruction under prepocessing to convert the data set into a new data set with a more convenient folder structure and file names and to convert DICOM images and segmentations into NIFTI files that are required as input to the automatic segmentation algorithm.

On TCIA, you can also download affine registration matrices (.tfm files) and the original segmentation contour lines (JSON files) under the "Version 2" tab.

Model

The implementation is based on MONAI and the network has a couple of differences compared to the network described in the paper:

• the inputs and outputs of the two convolution+PRelu+BatchNorm modules that are found in each layer are connected via a residual connection

• the downsampling is achieved by convolution with kernel_size=3 and stride=2, not by maxpooling.

• each convolution (except from the convolutions of the attention module) has a dropout layer following the PRelu activation

Requirements

The following setup has been tested on Ubuntu 20.04.

• A CUDA compatible GPU with memory not less than 4GB is recommended.

• CUDA Toolkit. Suggested version 10.2 or above.

• cuDNN. Suggested version 7.6.5 or above.

• Python. Suggested version is 3.6.

The following Python libraries can be installed individually, or with a single command:

    pip install -r requirements.txt

• PyTorch (recommended 1.6.0)

  pip install torch

• MONAI (recommended: 0.4.0)

  pip install monai

• natsort (recommended: 7.0.1)

  pip install natsort      

• matplotlib (recommended: 3.3.1)

  pip install matplotlib

• nibabel (recommended: 3.3.1)

  pip install nibabel

• torchvision (recommended: 0.7.0)

  pip install torchvision

• tensorboard (recommended: 2.3.0)

  pip install tensorboard

How to store/access the data

After downloading the repository VS_Seg, create a folder VS_Seg/data and place your data in it, e.g. the folder VS_defaced which contains the data as shown in the figure below.

In case, the folder structure or file names are different from the suggested one, or to exclude cases follow these instructions:

Open VS_Seg/Params/VSparams.py and make sure the function load_T1_or_T2_data returns train_files, val_files, test_fileswhere each of the returned variables is a list of dictionaries. Each dictionary represents an image/label pair, where the key "image" corresponds to the path to the image file and the key "label" corresponds to the path to the label file. train_files contains all dictionaries of paths to the training images/labels, val_filesto the validation images/labels and test_files to the testing images/labels.

How to train

To start training, run the following command from a terminal in the VS_Seg repository:

    python3 VS_train.py --results_folder_name $RESULTS_FOLDER_NAME 2> train_error_log.txt

optional parameters:

--results_folder_name followed by path to folder in which results are to be saved

--split followed by path to CSV file. See params/split_TCIA.csv as an example.

--no_attention removes attention module from neural network and attention maps from loss function

--no_hardness removes voxel hardness weighting from loss function

--dataset T1 or --dataset T2 to select the T1 or T2 dataset (default: T1)

--debug to run the algorithm on a small dataset, for only a couple of epochs

This script will create the folder VS_crop/model/results/ and the three sub-folders

• figures ⟶ will contain after training is completed:

  • a png file of a central slice (center of mass slice of label) of the first training set image after all pre-processing transformations have been applied to it, i.e. the way it is passed into the neural network, and the corresponding label
  • a png file showing training curves mean loss vs. epoch number and mean validation dice score vs. epoch number

• logs ⟶ contains the full training log (training_log.txt). Can be used to track training process during training.

• model ⟶ stores the model that performed best on the validation set (best_metric_model.pth)

• inferred_segmentations_nifti ⟶ contains the segmentations from inference (see next section)

Additionally, all output and error messages are logged in VS_Seg/train_error_log.txt

All model and optimizer parameters can be changed in the __init__ function of class VSparams in VS_Seg/Params/VSparams.py

How to run inference

To start the training run the following command from a terminal in the VS_Seg repository:

    python3 VS_inference.py --results_folder_name $RESULTS_FOLDER_NAME 2> inference_error_log.txt