An imaging-based method of mapping multi-echo BOLD intracranial pulsatility.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiac-related intracranial pulsatility is linked to cerebrovascular health.
• Blood-oxygen-level-dependent (BOLD) Magnetic Resonance Imaging (MRI) data contains information on intracranial pulsatility.
• Accurate isolation of BOLD pulsatility is of significant interest for understanding brain health.

Purpose of the Study:

• To investigate a deep learning approach for isolating and synthesizing BOLD pulsatility maps.
• To evaluate the performance of a U-Net model in characterizing BOLD physiological signals.

Main Methods:

• Multi-echo BOLD images, respiratory, and cardiac recordings were acquired from 55 adults.
• A U-Net deep learning model was trained using BOLD fast Fourier transform magnitude images as inputs.
• Model performance was assessed using metrics like Mean Squared Error (MSE), Mean Absolute Error (MAE), Structural Similarity Index (SSIM), and Mutual Information (MI).

Main Results:

• The U-Net model successfully synthesized BOLD pulsatility maps, with performance improving with more input images.
• While MAE showed no significant difference with or without the U-Net, SSIM and MI were significantly better with the U-Net.
• High cross-correlation (r > 0.90) was observed in the insula when comparing models with and without respiratory preprocessing.

Conclusions:

• A U-Net model can synthesize multi-echo BOLD pulsatility maps using temporal-frequency BOLD image inputs.
• This study contributes to the advancement of deep learning methods for characterizing physiological signals in BOLD MRI data.