DSA-Thrombus-Classifier
PoC demonstration tool to use CNNs to classify digital subtraction angiography images of the brain into thrombus-free and non-thrombus-free. User application which uses Mittmann et al.'s classifier. We describe this application in our paper Towards clinical translation of deep-learning based classification of DSA image sequences for stroke treatment.
This tool only represents a proof of concept and is no medical product!
Getting started
Before the first start, you need to setup the python service. For this, execute Backend/setup.bat . This might take a while.
In the meanwhile, you can download the pretrained model from our project page in the OpenScienceFramework, from the Demonstrator directory in the data section. Make sure to keep the structure of the folder.
You will also find a PDF version of our detailed installation manual there.
To use the application, start the python backend using Backend/RunServer.bat . Afterwards, open the frontend application using UI/Demonstrator.exe .
You first need to select the model you want to use. For this, select the directory of your models, in which the two subdirectories, frontal and lateral, are contained. Once this is selected, the application will remember this at every start until you change it again. Afterwards, you are ready to go. For this, select the frontal and lateral image respectively. We support nifti as well as DICOM, however, for the ladder, you first need to convert the file using the conversion-button. Once both files are selected, the application displays a mid-sequence preview. You are then ready to run the classification using the respective robot-button.
You might want to change the threshold that is used to determine whether a classification counts as thrombus-free or non-thrombus-free. By default, we recommend using a threshold of 0.57, which is done by optimizing a trade-off between the different metrics.
Settings
Before the first use, you have to setup your appsettings.json file.
| Option | Description |
|---|---|
AiServiceUrl |
The address and port where the python service serves at. When deploying the python service on the same machine, the default address is fine. |
PlastimatchPath |
The path to the plastimatch executable. If no valid path is given (default), the application will look in the working directory for the executable. |
ModelsEvaluationDirectory |
The path to the test results which are used to calculate the metrics. By default, the application will look in the in the working directory for a folder named TestResults. |
EnableEvaluationSetup |
Use this option to log the outputs of the model. |
LogPath |
In case EnableEvaluationSetup is set to true, use this option to specify the folder where the results are written to. By default, the working directory will be used. |
Important: When shipping a new release, make sure that you do exclude the settings file! (Or tell the user how to migrate) Otherwise, the configuration will be lost.
Building
You need to include SimpleITKNative.dll and SimpleITKManaged.dll in the UI folder in order to be able to compile it.
Attribution
Icon made by Freepik from www.flaticon.com
Citation
If you like to use this tool or wish to refer to our work, please do so by citing:
@InProceedings{10.1007/978-3-658-41657-7_22,
author="Baumg{\"a}rtner, Timo
and Mittmann, Benjamin J.
and Malzacher, Till
and Ro{\ss}kopf, Johannes
and Braun, Michael
and Schmitz, Bernd
and Franz, Alfred M.",
editor="Deserno, Thomas M.
and Handels, Heinz
and Maier, Andreas
and Maier-Hein, Klaus
and Palm, Christoph
and Tolxdorff, Thomas",
title="Towards Clinical Translation of Deep Learning-based Classification of DSA Image Sequences for Stroke Treatment",
booktitle="Bildverarbeitung f{\"u}r die Medizin 2023",
year="2023",
publisher="Springer Fachmedien Wiesbaden",
address="Wiesbaden",
pages="95--101",
abstract="In the event of stroke, a catheter-guided procedure (thrombectomy) is used to remove blood clots. Feasibility of machine learning based automatic classifications for thrombus detection on digital substraction angiography (DSA) sequences has been demonstrated. It was however not used live in the clinic, yet. We present an open-source tool for automatic thrombus classification and test it on three selected clinical cases regarding functionality and classification runtime. With our trained model all large vessel occlusions in the M1 segment were correctly classified. One small remaining M3 thrombus was not detected. Runtime was in the range from 1 to 10 seconds depending on the used hardware. We conclude that our open-source software tool enables clinical staff to classify DSA sequences in (close to) realtime and can be used for further studies in clinics.",
isbn="978-3-658-41657-7"
}