Skip to content

AutoAIF Integration

DCEPrep uses AutoAIF (formerly vascular_function), a companion repository from the petmri group, to perform automated Arterial Input Function (AIF) detection via a pre-trained deep learning model.

Note

This project was renamed from vascular_function to AutoAIF and moved to github.com/petmri/AutoAIF. Existing installs cloned from the old repository should re-clone from the new location.

What AutoAIF Does in DCEPrep

The AIF describes the concentration of contrast agent in blood over time and is required for pharmacokinetic modeling. Selecting the AIF manually from DCE-MRI data is time-consuming and operator-dependent. AutoAIF provides a neural network (main_vif.py) that identifies arterial voxels automatically from the DCE image.

DCEPrep calls AutoAIF during Preprocessing Step 12 when the -A flag is passed to preprocess_all.sh.

Mode Flag Behavior
Automatic -A A Neural network selects AIF; no manual input needed
Manual-preferred -A M Uses manual mask if present, falls back to automatic
Training -A T Uses manual mask and generates training data for the model

Outputs

AutoAIF produces the following files per subject:

File Description
*_desc-AIFfloat_mask.nii Continuous probability mask of arterial voxels
*_desc-AIFtopvoxels_mask.nii Thresholded binary mask (top voxels selected)
*_desc-AIF_resampledcurve.svg Plot of the resampled AIF curve
*_desc-AIF_mask.svg Visualization of the AIF mask overlaid on the image

How DCEPrep Locates AutoAIF

At runtime DCEPrep searches for main_vif.py:

AUTO_AIF_PATH=$(find $HOME -wholename '*main_vif.py' -printf '%h\n' -quit \
  || find / -name '*main_vif.py' -printf '%h\n' -quit)

In Docker, AutoAIF is installed at /opt/vascular_function and the pre-trained model weights are mounted at:

/opt/vascular_function/docker/files/model_weight_huber1.h5

Installing AutoAIF Without Docker

git clone https://github.com/petmri/AutoAIF.git
pip install -r AutoAIF/requirements.txt

Note

The Docker installation filters out cupy and tensorrt_cu12 to avoid GPU library conflicts. If installing manually in a GPU environment, review the requirements file and adjust CUDA-dependent packages for your setup.

Ensure main_vif.py is discoverable from your $HOME directory, or set AUTO_AIF_PATH manually before running preprocess_all.sh.