Validation of deep-learning accelerated quantitative susceptibility mapping for deep brain nuclei.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Quantitative Susceptibility Mapping (QSM) is crucial for neuroimaging, but prolonged scan times limit its clinical application.
• Deep brain nuclei evaluation using QSM is essential for understanding various neurological conditions.

Purpose of the Study:

• To assess the feasibility and consistency of a deep-learning (DL) accelerated QSM method for deep brain nuclei.
• To compare DL-QSM with conventional parallel imaging (PI)-QSM under different acceleration factors (AF).

Main Methods:

• Developed a DL-QSM method using a cascaded Convolutional Neural Network (CNN) with a Poisson disk undersampling scheme.
• Acquired QSM data from 59 participants using both PI-QSM (AF=2) and DL-QSM (AF=3, 4, 5) with identical imaging parameters.
• Evaluated image similarity using Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR), and assessed QSM value agreement and age-QSM correlations in 7 deep brain nuclei.

Main Results:

• DL-QSM images showed high similarity to PI-QSM images (mean SSIM: 0.85-0.87).
• Mean PSNR values for DL-QSM were consistently high (44.23-44.56).
• Susceptibility values from DL-QSM were within 5% of PI-QSM values at the group level, and age-QSM associations were consistently observed.

Conclusions:

• DL-accelerated QSM is a viable method for measuring susceptibility in deep brain nuclei.
• Acceleration factors up to 5 in DL-QSM do not significantly alter the age-susceptibility relationship in deep brain nuclei, enabling faster scans.