TV-BASED DEEP 3D SELF SUPER-RESOLUTION FOR FMRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Functional Magnetic Resonance Imaging (fMRI) provides insights into cognitive processes but has spatial resolution limitations.
• Current Deep Learning (DL) Super-Resolution (SR) methods for fMRI often require ground truth (GT) high-resolution (HR) data, which is difficult to obtain.
• Existing SR techniques face limitations in enhancing fMRI resolution due to data acquisition constraints and the trade-offs between temporal resolution, spatial resolution, signal-to-noise ratio, and scan time.

Purpose of the Study:

• To develop a novel self-supervised DL SR model for enhancing fMRI resolution.
• To overcome the dependency on GT HR data in DL-based fMRI SR.
• To improve the fine-grained analysis of brain functional architecture by increasing fMRI spatial resolution.

Main Methods:

• Introduced a self-supervised DL SR model combining a DL network with an analytical approach.
• Incorporated Total Variation (TV) regularization into the SR model.
• Eliminated the need for external GT HR images during the training process.

Main Results:

• The proposed self-supervised DL SR model achieved competitive performance compared to supervised DL techniques.
• The method successfully generated high-resolution (HR) fMRI images from low-resolution (LR) images.
• Functional maps were preserved during the super-resolution process.

Conclusions:

• Self-supervised DL SR offers a viable alternative to supervised methods for enhancing fMRI resolution.
• The novel approach removes the bottleneck of acquiring GT HR data, making fMRI SR more accessible.
• This method holds promise for advancing the detailed analysis of brain functional architecture in neuroscience research.