High resolution multi-delay arterial spin labeling with self-supervised deep learning denoising for pediatric choroid...

Area of Science:

• Neuroimaging
• Pediatric Radiology
• Artificial Intelligence in Medicine

Background:

• The choroid plexus (CP) produces cerebrospinal fluid (CSF), crucial for brain function.
• Multi-delay arterial spin labeling (MD-ASL) assesses CP perfusion but is challenging in children due to small CP size.
• Previous studies have not optimized MD-ASL for pediatric populations.

Purpose of the Study:

• To develop and validate a high-resolution (iso2 mm) MD-ASL protocol for pediatric CP perfusion assessment.
• To introduce a Transformer-based deep learning (DL) model for denoising MD-ASL images in children.
• To evaluate the DL model's performance against benchmark methods for improved SNR and repeatability.

Main Methods:

• A 10-minute high-resolution (iso2 mm) MD-ASL protocol was applied to 21 typically developing children (aged 8-17).
• A Transformer-based DL model was trained using k-space weighted image average (KWIA) denoised images.
• Performance was assessed via Signal-to-Noise Ratio (SNR), bias, and repeatability of perfusion parameters in CP and gray matter.

Main Results:

• The Transformer DL model effectively denoised pediatric MD-ASL images, significantly improving SNR for individual delays and fitted perfusion maps.
• The proposed method demonstrated superior performance compared to KWIA, TGV regularization, and Noise2Void.
• Enhanced SNR facilitates better visualization and quantification of CP perfusion in children.

Conclusions:

• The developed high-resolution MD-ASL protocol and Transformer DL model enable accurate CP perfusion assessment in children.
• This method holds potential for neurodevelopmental studies, aiding in characterizing CP and glymphatic system development.
• Facilitates advanced pediatric neuroimaging research and clinical applications.