Deep Learning Interactive Visualization

This project contains all code to learn a convolutional neural network model to detect Alzheimer's disease and visualize contributing brain regions with high relevance.

Further details on the procedures including samples, image processing, neural network modeling, evaluation, and validation were published in:

Dyrba et al. (2021) Improving 3D convolutional neural network comprehensibility via interactive visualization of relevance maps: evaluation in Alzheimer’s disease. Alzheimer's research & therapy 13. DOI: 10.1186/s13195-021-00924-2.

Screenshot of the InteractiveVis app

Running the interactive visualization

The interactive Bokeh web application InteractiveVis can be used for deriving and inspecting the relevance maps overlaid on the original input images.

To run it, there are three options.

1. We set up a public web service to quickly try it out: https://explaination.net/demo

2. Alternatively, download the docker container from DockerHub: sudo docker pull martindyrba/interactivevis Then use the scripts sudo ./run_docker_intvis.sh and sudo ./stop_docker_intvis.sh to run or stop the Bokeh app. (You find both files above in this repository.) After starting the docker container, the app will be available from your web browser: http://localhost:5006/InteractiveVis

3. Download this Git repository. Install the required Python modules (see below). Then point the Anaconda prompt or terminal console to the DeepLearningInteractiveVis main directory and run the Bokeh app using: bokeh serve InteractiveVis --show

Requirements and installation:

To be able to run the interactive visualization from the Git sources, you will need Python <3.8, in order to install tensorflow==1.15. Also, we recommend to first create a new Python environment (using Anaconda or virtualenv/venv) to avoid messing up your local Python modules/versions when you have other coding projects or a system shared by multiple users. ```console

for Anaconda:

conda create -n InteractiveVis python=3.7 conda activate InteractiveVis ```

Run pip to install the dependencies: console pip install -r requirements.txt

Then you can start the Bokeh application: console bokeh serve InteractiveVis --show

CNN model training and performance evaluation

The code for training the CNN models and evaluation is provided in this repository in the subdirectory scripts. The order of script execution was as follows:

• 1CreateResidualsADNI2_StoreModels.ipynb and other scripts for the validation samples 4CreateResidualsDELCODEapplyingADNI2regrmodel.ipynb (execution time: each 15-30 minutes).
• 2Train3DCNNADNI2xValwbmwp1CAT12MNIshuffle_checkpoint.ipynb for model training based on tenfold cross-validation to evaluate general model accuracy for the residualized data (execution time: 2-10 hrs with CUDA-GPU) and 3Train3DCNNADNI2wholedatasetwbmwp1CAT12MNI_shuffle.ipynb for training the model based on the whole ADNI-GO/2 dataset.
• 5Validate3DCNNxValwbmwp1CAT12MNI_DELCODE.ipynb and 6Validate3DCNNwholedswbmwp1CAT12MNIDELCODE.ipynb for the evaluation of the models using the validation data sets (execution time: each 15-30 minutes with CUDA-GPU).
• 7Train3DCNNADNI2xValwbrawdatmwp1CAT12MNIshufflecheckpoint.ipynb and 8Train3DCNNADNI2wholedatasetwbrawdatmwp1CAT12MNIshuffle.ipynb for training the models for the raw datasets (execution time: each 2-10 hrs with CUDA-GPU).
• 9Validate3DCNNwholedswbrawdatmwp1CAT12MNI_DELCODE.ipynb and 9Validate3DCNNxValwbmwp1CAT12MNI_DELCODE.ipynb for the evaluation of the models using the validation data sets (execution time: each 15-30 minutes with CUDA-GPU).
• xextracthippocampusrelevancelrpCMP_DELCODE.ipynb to extract the hippocampus relevance for all models (execution time: 15-30 minutes with CUDA-GPU).
• xextractrelevancemapsasniftiDELCODE.ipynb to extract the relevance maps directly as nifti file for all participants/scans (execution time: 30 minutes with CUDA-GPU).
• hippocampusvolumerelevanceanalysisDELCODE.html for the baseline group separation analysis of hippocampus volume and the correlation analysis of hippocampus volume and relevance (see also other R/Rmd scripts).
• yocclusionanalysis.ipynb code for the occlusion sensitivity analysis (execution time: 90 minutes with CUDA-GPU).
• zCreateResidualsdemodatasetapplyingADNI2regr_model.ipynb to create the example files being used by the InteractiveVis demo. It contains a sample of 15 people per diagnostic group, representatively selected from the ADNI-2 phase based on the criteria: amyloid status (positive for Alzheimer's dementia and amnestic mild cogntive impairment, negative for controls), MRI field strength of 3 Tesla, RID greater than 4000, and age of 65 or older.

InteractiveVis architecture overview

InteractiveVis UML class diagram (v4)

Select subject UML sequence diagram (v3)

License:

Copyright (c) 2020 Martin Dyrba martin.dyrba@dzne.de, German Center for Neurodegenerative Diseases (DZNE), Rostock, Germany

This project and included source code is published under the MIT license. See LICENSE for details.