Sample Tasks

This document contains two sample tasks for the classification and segmentation pipelines.

The document will walk through the steps in Training Steps, but with specific examples for each task. Before trying to train these models, you should have followed steps to set up an environment and AzureML.

Sample classification task: Glaucoma Detection on OCT volumes

This example is based on the paper A feature agnostic approach for glaucoma detection in OCT volumes.

Downloading and preparing the glaucoma dataset

The dataset is available here ^[1].

After downloading and extracting the zip file, run the create_glaucoma_dataset_csv.py script on the extracted folder.

python create_dataset_csv.py /path/to/extracted/folder

This will convert the dataset to csv form and create a file dataset.csv.

Finally, upload this folder (with the images and dataset.csv) to Azure Blob Storage. For details on creating a storage account, see AzureML Setup. The dataset should go into a container called datasets, with a folder name of your choice (name_of_your_dataset_on_azure in the description below).

Creating the glaucoma model configuration and starting training

Next, you need to create a configuration file InnerEye/ML/configs/MyGlaucoma.py which extends the GlaucomaPublic class like this:

from InnerEye.ML.configs.classification.GlaucomaPublic import GlaucomaPublic
class MyGlaucomaModel(GlaucomaPublic):
    def __init__(self) -> None:
        super().__init__()
        self.azure_dataset_id="name_of_your_dataset_on_azure"

The value for self.azure_dataset_id should match the dataset upload location, called name_of_your_dataset_on_azure above.

Once that config is in place, you can start training in AzureML via

python InnerEye/ML/runner.py --model=MyGlaucomaModel --azureml

As an alternative to working with a fork of the repository, you can use InnerEye-DeepLearning via a submodule. Please check here for details.

Sample segmentation task: Segmentation of Lung CT

This example is based on the Lung CT Segmentation Challenge 2017 ^[2].

Downloading and preparing the lung dataset

Download and prepare DICOM data

The dataset ^[3][4] can be downloaded here.

You need to convert the dataset from DICOM-RT to NIFTI. Before this, place the downloaded dataset in another parent folder, which we will call datasets. This file structure is expected by the conversion tool:

├── datasets
│   ├── LCTSC
│   │   ├── LCTSC-Test-*
│   │   ├── LCTSC-Train-*
│   │   ├── LICENSE

Convert data to NIFTI

Next, use the InnerEye-CreateDataset commandline tools to create a NIFTI dataset from the downloaded DICOM files. After installing and building the tool, open a powershell terminal, navigate to the build folder containing the executable (e.g. .InnerEye-CreateDataset\Source\projects\InnerEye.CreateDataset.Runner\bin\x64\Debug\) and run the following (this may take a few minutes to complete):

.\InnerEye.CreateDataset.Runner.exe dataset --datasetRootDirectory=C:\Users\<User>\datasets --niftiDatasetDirectory=lung_nifti --dicomDatasetDirectory=LCTSC --geoNorm 1 1 3 --groundTruthDescendingPriority esophagus spinalcord lung_r lung_l heart

The --geonorm flag tells the tool to normalize the voxel sizes during conversion.
The --groundTruthDescendingPriority flag removes additional labels from voxels that have been labelled as more than one structure. This improves the model performance on this specific dataset as it contains some overlapping structures.

Now, you should have another folder under datasets containing the converted NIFTI files:

├── datasets
│   ├── LCTSC
│   │   ├── LCTSC-Test-*
│   │   ├── LCTSC-Train-*
│   │   ├── LICENSE
│   ├── lung_nifti
│   │   ├── 1
│   │   ├── 2
│   │   ├── 3
│   │   ├── ...
│   │   ├── dataset.csv
│   │   ├── info.txt

Upload Dataset to Azure

Finally, upload this folder (with the images and dataset.csv) to Azure Blob Storage. We recommened using the Azure Storage Explorer tool for this.

For details on creating a storage account, see AzureML Setup. All files should go into a folder in the datasets container, for example my_lung_dataset. This folder name will be passed into InnerEye using the azure_dataset_id flag of the model configuration (see below).

Creating the lung model configuration and starting training

You can then create a new model configuration, based on the template Lung.py. To do this, create a file InnerEye/ML/configs/segmentation/MyLungModel.py, where you create a subclass of the template Lung model, and add the azure_dataset_id field (i.e., the name of the folder that contains the uploaded data from above), so that it looks like:

from InnerEye.ML.configs.segmentation.Lung import Lung
class MyLungModel(Lung):
    def __init__(self) -> None:
        super().__init__()
        self.azure_dataset_id = "my_lung_dataset"

If you are using InnerEye as a submodule, please add this configuration in your private configuration folder, as described for the Glaucoma model here.

You can now run the following command to start a job on AzureML:

python InnerEye/ML/runner.py --azureml --model=MyLungModel

See Model Training for details on training outputs, resuming training, testing models and model ensembles.

References

[1] Ishikawa, Hiroshi. (2018). OCT volumes for glaucoma detection (Version 1.0.0) [Data set]. Zenodo. http://doi.org/10.5281/zenodo.1481223

[2] Yang, J. , Veeraraghavan, H. , Armato, S. G., Farahani, K. , Kirby, J. S., Kalpathy-Kramer, J. , van Elmpt, W. , Dekker, A. , Han, X. , Feng, X. , Aljabar, P. , Oliveira, B. , van der Heyden, B. , Zamdborg, L. , Lam, D. , Gooding, M. and Sharp, G. C. (2018), Autosegmentation for thoracic radiation treatment planning: A grand challenge at AAPM 2017. Med. Phys.. . doi:10.1002/mp.13141

[3] Yang, Jinzhong; Sharp, Greg; Veeraraghavan, Harini ; van Elmpt, Wouter ; Dekker, Andre; Lustberg, Tim; Gooding, Mark. (2017). Data from Lung CT Segmentation Challenge. The Cancer Imaging Archive. http://doi.org/10.7937/K9/TCIA.2017.3r3fvz08

[4] Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, Moore S, Phillips S, Maffitt D, Pringle M, Tarbox L, Prior F. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository, Journal of Digital Imaging, Volume 26, Number 6, December, 2013, pp 1045-1057. (paper)