ASL-QC

ASL-QC is a Python library for automated quality assessment of arterial spin labeling (ASL) cerebral blood flow (CBF) maps using deep learning. It is developed by the Detre Lab at the University of Pennsylvania and distributed as open-source software.

Overview

ASL-QC provides automated quality assessment of ASL-derived CBF maps. Given a single CBF NIfTI image, ASL-QC returns a continuous quality score between 0 and 1:

• 0 — lowest predicted quality
• 1 — highest predicted quality

The model operates directly on the CBF map and does not require the underlying ASL images, a structural image, or any additional input.

Installation

ASL-QC requires Python 3.13 or later. We recommend installing it in a dedicated environment, for example with conda:

conda create -n asl-qc python=3.13
conda activate asl-qc

ASL-QC can then be installed either from PyPI or directly from the source repository.

Option 1: Install from PyPI

python -m pip install asl-qc

Option 2: Install from repository

git clone https://github.com/xavibeltranurbano/ASL-QC
cd ASL-QC
python -m pip install .

Getting Started

Python API

from asl_qc import predict_quality

quality = predict_quality("/path/to/cbf.nii.gz")
print(quality)

By default the model runs on the GPU if one is available and on the CPU otherwise. To choose the device explicitly, pass device="cpu" or device="cuda":

quality = predict_quality("/path/to/cbf.nii.gz", device="cpu")

The returned quality score ranges from 0 to 1, with higher values indicating higher predicted CBF map quality.

Command-Line Interface

ASL-QC can also be run directly from the command line:

asl-qc --path-cbf /path/to/cbf.nii.gz

Use --device cpu or --device cuda to select the device (default: auto).

Input Preprocessing

Before quality estimation, each CBF image is automatically:

1. Converted to canonical RAS orientation
2. Resampled to 2.5 mm isotropic resolution
3. Cropped or padded around the nonzero CBF support to 96 × 96 × 64
4. Clipped to [-100, 100]
5. Scaled by 100

Training

The training scripts can be found in src/training_scripts/.

Model Development

ASL-QC was developed using a 3D residual convolutional neural network with four convolutional blocks and a fully connected regression head. Model development and evaluation used nested five-fold cross-validation, with folds stratified by scanner vendor and quality rating. The released model is an ensemble of the five models trained across the outer cross-validation folds. Each model was optimized using mean squared error (MSE) between the predicted quality score and the mean expert rating for the corresponding CBF map.

Training Dataset

The training dataset contains 450 CBF maps from multiple cohorts, sites, scanner vendors, ASL acquisition protocols, and background-suppression (BS) configurations.

Cohort	Scanner	Age (mean ± SD)	Female n (%)	PLD	Labeling time	Protocol	BS	Maps
ADNI	Siemens	73.3 ± 7.4	17 (40%)	1.9 s	0.7 s	2D PASL	No	43
ADNI	GE	74.9 ± 7.5	12 (40%)	2.025 s	1.45 s	3D PCASL	Yes	30
MESA	Siemens	66.9 ± 8.6	25 (56%)	2 s	1.5 s	3D PCASL	Yes	45
SPRINT	Siemens	65.5 ± 8.2	13 (42%)	1.5 s	1.5 s	2D PCASL	No	31
SPRINT	GE	74.2 ± 7.7	7 (23%)	2.025 s	1.45 s	3D PCASL	Yes	30
SPRINT	Philips	67.3 ± 8.9	21 (42%)	1.5 s	1.5 s	2D PCASL	No	50
DLBS	Philips	54.9 ± 22.6	16 (64%)	1.525 s	1.65 s	2D PCASL	No	25
CARDIA	Siemens	50.5 ± 3.5	12 (63%)	1.5 s	1.5 s	2D PCASL	No	19
NACC	Siemens	69.1 ± 12.8	29 (53%)	1.8 s	1.8 s	3D PCASL	Yes	55
VCID	Siemens	54.4 ± 9.2	26 (58%)	2 s	3 s	3D PCASL	Yes	45
VCID	GE	61.2 ± 11.0	22 (55%)	2 s	3 s	3D PCASL	Yes	40
Institutional data	Siemens	45.5 ± 19.2	18 (49%)	1.2–2 s	1.8–3 s	3D PCASL	Yes	37
Total	Siemens 275; GE 100; Philips 75	63.9 ± 14.4	218 (48%)	1.2–2.025 s	0.7–3 s	3D PCASL 282; 2D PCASL 125; 2D PASL 43	Yes 282; No 168	450

Rating Strategy

Four expert raters independently evaluated each CBF map using a five-point quality scale. The ground-truth quality score was defined as the mean of the four expert ratings and normalized to the range [0, 1].

Rating	Interpretation guidelines
0.00	No discernible brain anatomy; severe artifacts and/or poor resolution; completely uninterpretable.
0.25	Anatomy is barely recognizable; major artifacts and/or poor resolution; very limited interpretability.
0.50	Anatomy is recognizable; minor artifacts and/or smoothness; moderate spatial resolution; basic interpretability.
0.75	Good anatomical detail; limited artifacts and/or smoothness; good spatial resolution; good interpretability.
1.00	Excellent anatomical detail; no artifacts; high spatial resolution; good interpretability.

Citation

If you use ASL-QC or its pretrained model in your work, please cite:

Beltran-Urbano X, Taso M, Nasrallah I, Detre JA, Wang Z, Dolui S.
QEI-Net: A deep learning-based automated quality evaluation index for ASL CBF maps.
Proceedings of the 2025 ISMRM & ISMRT Annual Meeting & Exhibition. 2025;0730.

License

Please see the LICENSE file for licensing information.

Links

• PyPI