Automatic detection of arterial input function for brain DCE-MRI in multi-site cohorts

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate arterial input function (AIF) extraction is essential for quantitative pharmacokinetic modeling in dynamic contrast-enhanced MRI (DCE-MRI).
• Traditional AIF extraction relies on manual selection, which can be time-consuming and prone to inter-observer variability.
• Developing automated methods is crucial for improving the efficiency and reliability of DCE-MRI analysis.

Purpose of the Study:

• To develop and validate a robust deep learning model for precise arterial input function (AIF) extraction from DCE-MRI images.
• To assess the performance of the proposed model against manual AIF extraction using quantitative metrics.

Main Methods:

• A 3D UNet deep learning model was implemented and trained on manually delineated AIF regions from a diverse dataset of 384 human brain DCE-MRI scans across five institutions.
• The model's performance was evaluated using a novel AIF quality metric (AIFitness) and by comparing K^trans values derived from standard DCE pipelines.
• The validated model was then applied to a separate, independent dataset of 421 DCE-MRI scans for replication.

Main Results:

• The 3D UNet model achieved high AIFitness scores (93.9 on the primary test set, 97.0 on the replication set), closely approaching the performance of manual selections (99.7).
• Intraclass correlation for K^trans values between automated and manual AIFs was strong (0.89), indicating excellent agreement.
• Quantitative analysis showed no significant differences in white matter K^trans values between automated and manual AIF extraction methods.

Conclusions:

• The proposed 3D UNet model, incorporating enhanced convolutional kernels and a modified Huber loss function, demonstrates superior performance for automated AIF extraction from multi-center DCE-MRI data.
• The automated AIF extraction method provides reliable and consistent results comparable to manual selection, as evidenced by AIFitness scores and DCE-MRI-derived metrics like K^trans maps.
• This deep learning approach offers a promising solution for efficient and accurate pharmacokinetic analysis in DCE-MRI studies, particularly for blood-brain barrier permeability measurements.