We provide a tool for detection and segmentation of ischemic acute and sub-acute strokes in brain diffusion weighted MRIs (DWIs). The deep learning networks were trained and tested on a large dataset of 2,348 clinical images, and further tested on 280 images of an external dataset. Our proposed model outperformed generic nets and patch-wise approaches, particularly in small lesions, with lower false positive rate, balanced precision and sensitivity, and robustness to data perturbs (e.g., artefacts, low resolution, technical heterogeneity). The agreement with human delineation rivaled the inter-evaluator agreement; the automated lesion quantification (e.g., volume) had virtually total agreement with human quantification. The method has minimal computational requirements and is fast: the lesion inference takes 20~30 seconds in CPU and the total processing, including registration and generation of results/report takes ~ 2.5 mins. The outputs, provided with a single command line, are: the predicted lesion mask, the lesion mask and the image inputs (DWI, B0, ADC) in standard space, MNI, and the quantification of the lesion per brain structure and per vascular territory.
The latest version ADSv1 was uploaded on NITRC only for better maintenance. Please visit ADSv1 for more information. https://www.nitrc.org/frs/?group_id=1520#
The ${ROOT} is described as below.
${ROOT}
|-- data
|-- Trained_Nets
|-- examples
|-- Subject01
|-- Subject01
|-- template
|-- codes
datacontains data like templates, image examples, and trained networks.codescontains ADS pipeline bin codes and the main function.
- python (version 3.7.7): Please make sure the version is at least 3.6+
- tensorFlow (version 2.0.0): The Deep Learning networks library for backend.
- h5py (version 2.10.0) : The networks model storage library. (please make sure its version is 2.10.0. Any different version might generate unexpected error when loading our pretrained models.)
- niBabel (version 3.2.1): For loading NIFTI files.
- numpy (version 1.19.5): Gerenal computing array processing library.
- scipy (version 1.4.1): For image operation/processing.
- dipy (version 1.4.0): For image registration
- scikit-image (version 0.18.1): For image operation/processing.
- scikit-learn (version 0.24.1): Not necessary, but recommended because we will update codes with this dependency.
-
Cloned the codes (for unix system, similar steps should be sufficient for Windows) with :
git clone https://github.com/Chin-Fu-Liu/Acute-stroke_Detection_Segmentation/Download pre-trained networks :
- Download pre-trained networks from here (google drive)
- Unzip and put all the models ('MODEL_NAMES.h5' files) directly under
Trained_Netsfolder that is under yourAcute-stroke_Detection_Segmentationmain folder.
-
If you are not familiar with github, you can just download the whole ADS package (ADSv0.0.zip file) from google drive here (google drive) and unzip it to create the
Acute-stroke_Detection_Segmentationmain folder locally.
We highly recommend creating a virtual enviroment for using this software.
From a bash shell, create a virtual environment in a folder (FOLDER_FOR_ENVS/ENVS_FOLDER_NAME) that you want.
FOLDER_FOR_ENVS can be the path to the folder (Acute-stroke_Detection_Segmentation) you create and clone from github or google drive.
ENVS_FOLDER_NAME can be any name you like, ex: ADS_ENV.
Using Conda:
conda create -p FOLDER_FOR_ENVS/ENVS_FOLDER_NAME python=3.6.5 -y
source activate FOLDER_FOR_ENVS/ENVS_FOLDER_NAME
Using Virtualenv:
virtualenv -p python3 FOLDER_FOR_ENVS/ENVS_FOLDER_NAME # Use python up to 3.6+
source FOLDER_FOR_ENVS/ENVS_FOLDER_NAME/bin/activate
(Use pip install --user for local user)
$ pip install numpy nibabel scipy scikit-image scikit-learn
$ pip install dipy==1.4.0
$ pip install tensorflow==2.0.0
$ pip install tensorflow-gpu==2.0.0
$ pip install h5py==2.10.0
(We don't support GPU in this version yet, but it will be included in the next updated version. Hence, please also install tensorflow-gpu.)
Navigate to to the /Acute-stroke_Detection_Segmentation/codes folder, by cd PATH_TO_/Acute_Stroke_Detection/codes/
Under the Acute-stroke_Detection_Segmentation/codes folder, run
python ADSRun.py -input PATH_TO/SUBJECTID_FOLDER
The input images must be in Nifti format (.nii or .nii.gz). You can convert to Nifti using any available software/script (e.g., dcm2nii, MRICron, ImageJ).
SUBJECTID_FOLDER should be named by its SubjectID (see our example under "data/examples"). The data storage structure and naming must be as follows:
|-- SUBJECTID_FOLDER
|-- SUBJECTID_DWI.nii.gz
|-- SUBJECTID_b0.nii.gz
|-- SUBJECTID_ADC.nii.gz (optional)
The mandatory inputs are DWI and B0; ADC is optional. If no ADC is provided, it will be calculated with the b-value provided by the user (see option -bvalue). If no b-value is provided, it will use the default, b=1000.
For detail description, run -h for help as following
python ADSRun.py -h
-input is the path for SUBJECTID_FOLDER
-model is the model name for segmenting lesions. It can be DAGMNet_CH3, DAGMNet_CH2, UNet_CH3, UNet_CH2, FCN_CH3, and FCN_CH2. These models were pretrianed by our data, as reported in our paper [cite:] (defaut: DAGMNet_CH3)
-bvalue is used to specify the b-value to calculate ADC, if ADC is not given. If ADC is given under SUBJECTID_FOLDER, this option will be ignored. (defaut: 1000)
-save_MNI is used to specify whether to save images in MNI space (DWI, b0, ADC, Normalized DWI and lesion predict). You can turn it off as -save_MNI False. (defaut: True)
-generate_brainmask is used to specify whether to generate brain mask in raw space, brain mask in MNI will be generated if -save_MNI is True. You can turn it on as -save_MNI True. (defaut: False)
-generate_report is used to specify whether to generate the “lesion report”. The lesion report shows the total lesion volume and the total intracranial volume in the subject's original space, as well as the estimated number of lesion voxels per brain structure and per vascular territory, in MNI space. The folder "data/template" contains the eletronic atlases defining these areas as well a list of their volumes, in the MNI template. Therefore, users can calculate the relative distribution of the lesions (e.g., the percentage of a given vascular territory affected by the stroke). Be aware that these values are calculated by linear mapping to MNI space, therefore the precision highly depends on the particular brain morphology. You can turn it off as -generate_report False. (defaut: True)
-generate_result_png is used to specify whether to generate a figure (.png) of DWI, b0, ADC, and DWI aligned with lesion predict (blue contour) in the original image space. This figure is useful for immediate quality checking. You can turn it off as -generate_result_png False. (defaut: True)
For example, if you want to get a lesion predict on Subject01 with DAGMNet_CH3 model, as well as the images in MNI space, the lesion report, and the figure for quality control, you simply run the following line in your virtual environment under the Acute_Stroke_Detection/codes folder.
python ADSRun.py -input PATH_TO_Subject01_FOLDER
if you want to get a lesion predict on Subject01 with UNet_CH2 model, but not generating images in MNI and lesion report, you can run the following code in your virtual environment under the Acute_Stroke_Detection/codes folder.
python ADSRun.py -input PATH_TO_Subject01_FOLDER
-model UNet_CH2
-save_MNI False
-generate_report False
-generate_result_png True
If you are not running the code under under the Acute_Stroke_Detection/codes folder, then you need to specify the path to Acute_Stroke_Detection/codes folder, as follows:
python PATHTO/Acute_Stroke_Detection/codes/ADSRun.py -input PATH_TO_Subject01_FOLDER
- 2021.04.22. ADSv0.0 was uploaded.
For stroke segmentaion and quantification: Liu CF, Hsu J, Xu X, Ramachandran S, Wang V, Miller MI, Hillis AE, Faria AV. Deep learning-based detection and segmentation of diffusion abnormalities in acute ischemic stroke. Commun Med 1, 61 (2021) https://doi.org/10.1038/s43856-021-00062-8
For ASPECTS calculation: Liu CF, Li J, Kim G, Miller MI, Hillis AE, Faria AV. Automatic comprehensive aspects reports in clinical acute stroke MRIs. Sci Rep 13, 3784 (2023). https://doi.org/10.1038/s41598-023-30242-6
For automatic radiological reports: Liu CF, Zhao Y, Yedavalli V, Leigh R, Falcao V, Miller MI, Hillis AE, Faria AV. Automatic comprehensive radiological reports for clinical acute stroke MRIs. Commun Med 3, 95 (2023). https://doi.org/10.1038/s43856-023-00327-4
For the dataset: Liu CF, Leigh R, Johnson B, Urrutia V, Hsu J, Xu X, Li X, Mori S, Hillis AE, Faria AV. A large public dataset of annotated clinical MRIs and metadata of patients with acute stroke. Sci Data 10, 548 (2023). https://doi.org/10.1038/s41597-023-02457-9
For the dataset source: Faria, Andreia V. Annotated Clinical MRIs and Linked Metadata of Patients with Acute Stroke, Baltimore, Maryland, 2009-2019. Inter-university Consortium for Political and Social Research [distributor], 2022-12-12. https://doi.org/10.3886/ICPSR38464.v5
This work is licensed under GNU General Public License v3.0, as found in the LICENSE file.