Improving Arterial Spin Labeling by Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Arterial Spin Labeling (ASL) is a non-invasive MRI technique for quantifying cerebral blood flow.
• Conventional ASL requires multiple subtraction images, increasing scan time and susceptibility to artifacts.
• Developing faster and more robust ASL methods is crucial for clinical applications.

Purpose of the Study:

• To develop a deep learning algorithm for generating accurate and robust ASL perfusion images.
• To reduce the number of required subtraction images for ASL generation.
• To improve the signal-to-noise ratio and reduce artifacts in ASL imaging.

Main Methods:

• A convolutional neural network (CNN) was employed for ASL image generation from pairwise subtraction images.
• Ground truth perfusion images were generated by averaging 6-7 subtraction images from healthy subjects and patients.
• CNNs were trained using 2-3 subtraction images and validated through cross-validation and testing on stroke datasets.

Main Results:

• CNNs achieved approximately 40% lower mean square errors compared to the conventional averaging method.
• Region-of-interest analysis in stroke regions showed significantly smaller mean square errors for CNN-generated cerebral blood flow maps.
• Radiologic scoring indicated superior suppression of noise, motion, and segmentation artifacts by CNNs.

Conclusions:

• CNNs provide superior perfusion image quality in ASL compared to conventional averaging methods.
• Deep learning enhances the accuracy of ASL perfusion measurements.
• This approach enables robust ASL image generation using a reduced number of subtraction images.