Workflow for automatic renal perfusion quantification using ASL-MRI and machine learning.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Renal arterial spin labeling (ASL) MRI is crucial for assessing kidney perfusion.
• Clinical use of renal ASL MRI is limited by manual segmentation, which is time-consuming and observer-dependent.
• Machine learning (ML) shows promise for automating medical image segmentation, including kidneys.

Purpose of the Study:

• To develop and validate a fully automatic workflow for renal cortex perfusion quantification using ML-based segmentation.
• To replace manual segmentation in ASL quantification with an automated process.
• To enhance the clinical applicability of renal ASL MRI.

Main Methods:

• A cascade of three U-net models was developed for automatic renal cortex segmentation.
• The ML models were trained and validated using 1.5T ASL-MRI data from healthy volunteers.
• Cortical renal blood flow (RBF) values from automated segmentations were compared with manual segmentations.

Main Results:

• The automated segmentation achieved good agreement with manual segmentations (Dice score 0.78 ± 0.04 on training, 0.75 ± 0.03 on validation).
• Automated segmentation performance was comparable to inter-observer variability (Dice score 0.77 ± 0.02).
• Automated and manual methods yielded similar cortical RBF values (211 ± 31 vs. 208 ± 31 mL/min/100 g), with narrow limits of agreement.

Conclusions:

• The proposed ML-based workflow automates renal ASL quantification without sacrificing RBF accuracy.
• This automated approach makes renal ASL-MRI more clinically attractive, efficient, and suitable for longitudinal and multi-center studies.
• The method offers quick processing and eliminates observer dependence, facilitating wider adoption of renal ASL MRI.