AIF Selection¶

The Arterial Input Function (AIF) describes the contrast agent concentration in blood plasma over time and is required for quantitative Ktrans estimation. DCEPrep supports manual, semi-automatic, and fully automatic AIF selection.

Overview¶

The AIF is represented as a binary mask identifying voxels within the DCE image that correspond to arterial signal. During kinetic modeling, the mean signal within the AIF mask is used as the input function.

DCEPrep enforces that all AIF voxels fall within the brain mask, preventing contamination from extracranial vessels.

Selection Modes¶

The -A flag controls AIF selection mode in preprocess_all.sh:

Flag	Mode	Description
(omitted)	Manual	Use a pre-existing AIF mask file (must be placed at the expected BIDS path)
`-A A`	Automatic	Neural network selects AIF voxels automatically
`-A M`	Manual-preferred	Use manual mask if found; fall back to automatic
`-A T`	Training	Use manual mask and generate training data for the neural network

Automatic AIF (Neural Network)¶

DCEPrep's automated AIF detection uses a neural network from the companion AutoAIF repository (formerly vascular_function).

Requirements¶

The AutoAIF repo must be available (cloned or mounted in Docker)
A pre-trained weights file is required; specify its path with -w [path]
In Docker, the default weights file is located at /opt/vascular_function/model_weight_huber1.h5

How it works¶

The network is invoked via main_vif.py from the AutoAIF repository. It takes the 4D DCE time-series as input and produces a continuous probability mask indicating the likelihood that each voxel belongs to an artery. The probability mask is then thresholded to produce a binary AIF mask.

The selected AIF voxels are intersected with the brain mask so that only intracranial arterial voxels are included.

Outputs¶

File	Description
`desc-AIF_mask.nii.gz`	Binary AIF mask
`desc-AIF_float.nii.gz`	Continuous probability mask (pre-threshold)
`desc-AIF_mask.svg`	Visualization of the selected AIF region
`desc-AIF_curve.svg`	Plot of the mean AIF signal over time

Training mode (`-A T`)¶

Training mode processes the data using the manual AIF mask but also runs the network to generate prediction outputs. This produces paired manual/automatic data that can be used to retrain or fine-tune the network weights.

AIF Quality Metrics¶

The quality of the selected AIF is scored by aif_metric.py using four weighted metrics combined into an AIFitness score (0–100 scale):

Metric	Weight	Description
Peak ratio	30%	Sigmoid-weighted ratio of AIF peak to mean signal — higher peaks indicate better arterial contrast
Tail ratio	30%	Washout quality — how well the signal returns toward baseline after the peak
Baseline-to-mean	30%	Baseline stability — a low pre-contrast baseline relative to the mean indicates good contrast uptake
Peak timing	10%	Whether the peak occurs early in the acquisition, as expected physiologically

The AIFitness score is displayed in the case report. Cases with AIFitness below 59 are automatically flagged in the population report.

Manual AIF¶

To provide a manual AIF mask:

Draw the mask using an image viewer that supports NIfTI editing (e.g., FSLeyes, ITK-SNAP, or 3D Slicer). Select voxels within a major intracranial artery (typically the internal carotid or middle cerebral artery) on the DCE reference volume.
Save the mask as a binary NIfTI file (voxel values of 0 or 1) in the same space and dimensions as the DCE image.
Place the file at the expected BIDS path within the derivatives folder:

derivatives/dceprep/sub-##/ses-##/dce/sub-##_ses-##_desc-AIF_mask.nii.gz

A custom suffix can be specified with the -a [suffix] flag if your mask uses a different naming convention (.nii.gz is appended automatically).

AIF File Naming¶

The AIF mask file is expected at:

dce/sub-##_ses-##_desc-AIF_mask.nii.gz

A custom suffix can be specified with the -a [suffix] flag (.nii.gz is appended automatically).