Submillimeter diffusion MRI using an in-plane segmented 3D multi-slab acquisition and denoiser-regularized reconstruction

• Whole-cerebrum, in-vivo 3D diffusion MRI at 0.53–0.65 mm isotropic resolutions. • In-plane segmented 3D multi-slab EPI for SNR-efficient submillimeter diffusion MRI. • Denoiser-regularized reconstruction enhances SNR while maintaining image fidelity. • Submillimeter data reduce gyral bias and improve U-fiber mapping. • Robust performance across 3T and 7T, with scanner-agnostic Pulseq implementation. Diffusion MRI (dMRI) enables brain connectivity mapping but is constrained by spatial resolution. Previous post-mortem studies have demonstrated the potential of submillimeter dMRI in enabling more precise delineations of curved and crossing white matter pathways. However, achieving such resolution in-vivo poses significant challenges due to the intrinsically low signal-to-noise ratio (SNR). Furthermore, for echo-planar imaging (EPI), large matrix sizes often require long echo spacing, readout duration, and echo times (TE), leading to significant image distortion, T2* blurring, and T2 signal decay. Here, we propose an acquisition and reconstruction framework to overcome these challenges. Based on numerical simulations, we employ in-plane segmented 3D multi-slab EPI that leverages the optimal SNR efficiency of 3D multi-slab imaging while reducing echo spacing, readout durations, and TE using in-plane segmentation. This approach minimizes distortion, improves image sharpness, and enhances SNR. Additionally, we develop a denoiser-regularized reconstruction to suppress noise while maintaining data fidelity, which reconstructs high-SNR images without introducing substantial blurring or bias. At 3T, we present 0.53–0.65 mm in-vivo data that reveal finer fiber architectures, reduced gyral bias, and improved U-fiber mapping compared to 1.22 mm data. At 7T, we acquire 0.61 mm data that show excellent agreement with high-resolution post-mortem dMRI, demonstrating robustness and high SNR at an ultra-high field. Our method is implemented using the open-source, scanner-agnostic framework Pulseq to facilitate broader adoption across scanner platforms to benefit a wider range of applications. These results establish our approach as a promising tool for high-resolution dMRI, advancing neuroanatomical investigations of white matter architecture.