Hypothalamus segmentation

This repository provides the source code and models for segmentation of the hypothalamus in T1-weighted MR images in coronal orientation at resolution of 0.125x0.125x0.5 mm3 and intracranial volume (ICV) from coronal MRI images of resolution 0.5x0.5x0.5 mm3. The method is based on a convolutional neural network EfficientNetB0, which outputs segmentations of the hypothalamus and ICV.

Training dataset

During labeling of the data for each patient a folder containing subfolders for each single imaging slice is created. Two files are located in each subfolder: mri_orig.png for original MRI image and mri_hypothal.png for the manually created segmentation mask. All folders with labeled data are than put into folder named DATA which is on the same directory level as the repository. With generate_data.py script, the lookup table for training and validation (explicit names are set in the validation_folders variable within the script) data is created. As a result, the script creates the folder metadata_training with the dataset files shown in the structure below. Subject folders used for testing should not be stored in the DATA folder, since they should not be used in the training. The training data for ICV are created and stored analogically.

The Figure below illustrates hypothalamus training samples as overlays of the manually segmented masks on high-resolution MRI images.

Project structure

 
hypothalamus_segmentation
│   README.md
│   network.py
|   generate_data.py
│   my_generator.py
|
└───augmentation
│   │   extend_folder.py
│   │   adjust_contrast.py 
|
└───metadata_training
│   │   partition.npy
│   │   data_set.npy
|
└───validation
│   │   predict.py
|   |   efficientnetb0ICV.h5
|   |   efficientnetb0.h5
│   │   convert2npz.py
|   |   test_folder
|   |   test_folder.npz
|
└───overlay_images
│   │   GT
│   │   MRI
│   │   overlay
│   │   overlay_GT
│   │   overlayMRI
│   │   pred
│   │   overlayGT.py
│   │   overlayMRI.py
|
DATA
└───Folder 1
    | subfolder 1
    ...
    | subfolder m
...
└───Folder n

Data augmentation

extend_folder.py script should be used to respectively increased the number of new folders into which augmented files will be stored. Than adjust_contrast.py is used for contrast shifting with some sampled values and creation of new training data.

Training

Training can be started by executing network.py file. my_generator.py contains the code that is used by the lookup table to load the training samples from the dataset and passes them to the network.

Ouput models

can be downloaded from Releases

• models.zip/efficientnetb0 for hypothalamus segmentation
• models.zip/efficientnetb0ICV for intracranial volume segmentation

Testing

convert2npz.py script is used to create test dataset test_folder.npz from the subject folder test_folder with the same structure as the training dataset. Then predict.py script can be executed for prediction and storing of the files pridected.png in the same subfolders as the original test data and calculation of different metrics and number of pixels for each single subject.

The Figure below shows two segmentation examples (in coronal slices) obtained for a control subject and a subject diagnosed with Amyothrophic Lateral Sclerosis (ALS) disease.

Additional libraries:

• Segmentation Models
• Surface distance metrics

must be installed via:

 pip install <package name>

The Figure below illustrates automatically segmented ICV overlayed on original MRI images.