Ultra-High Resolution TR-external EPIK with Deep Learning Image Reconstruction for Enhanced Characterisation of Cortical Depth-dependent Neural Activity

The detection of neural signals using functional MRI at the laminar or columnar level enables non-invasive exploration of fundamental brain function processing and the interconnected pathways within intracortical tissues. The growing interest in this research area is driven by advancements in fMRI acquisition techniques that enhance spatial resolution for high-fidelity mapping. However, the submillimetre voxel sizes commonly employed in layer-specific fMRI studies raise concerns about relatively low signal-to-noise ratios and increased image artefacts in reconstructed images, ultimately limiting the precise delineation of cortical depth-dependent functional activities.

This work aims to address this issue by incorporating a deep learning technique for enhanced image reconstruction of the submillimetre fMRI data, acquired with echo-planar-imaging with keyhole (EPIK) combined with the repetition-time-external (TR-external) EPI phase correction scheme. Our network was trained in a self-supervised, scan-specific manner using the sampling strategy from zero-shot self-supervised learning (ZS-SSL) method, which has gained attention for high-resolution MR image reconstruction.

Healthy volunteers participated in this study, and the performance of the developed method was evaluated in direct comparison to the conventional reconstruction method using datasets acquired at 7T. The deep learning reconstruction produced reconstructed images with significantly higher SNR than the conventional method, which was further quantitatively validated through histogram analysis. This enhancement was consistent across all slice locations, demonstrating the reliability of the scan-specific deep learning technique.