Multi-contrast MRI and histology datasets used to train and validate MRH networks to generate virtual mouse brain histology

H MRI maps brain structure and function non-invasively through versatile contrasts that exploit inhomogeneity in tissue micro-environments. Inferring histopathological information from MRI findings, however, remains challenging due to absence of direct links between MRI signals and cellular structures. Here, we provided deep convolutional neural networks, called MRH-Nets, developed using co-registered multi-contrast MRI and histological data of the mouse brain, can estimate histological staining intensity directly from MRI signals at each voxel. The results provide three-dimensional maps of axons and myelin with tissue contrasts that closely mimics target histology and enhanced sensitivity and specificity compared to conventional MRI markers. The dataset contains multi-contrast MRI and histology used for the training and testing and the acquisition parameters. The datasets have been carefully registered to mouse brain images from the Allen Mouse Brain Atlas (https://mouse.brain-map.org). The source codes for MRH-Nets can be found at https://github.com/liangzifei/MRH-Net.

Methods
Ex vivo Mouse brain MRI: Adult C57BL/6 mice (P60, n=10, Charles River, Wilmington, MA, USA), rag2-/- shiverer and littermate controls (n=5/5, P50) were perfusion fixed with 4% paraformaldehyde (PFA) in PBS. The samples were preserved in 4% PFA for 24 hours before transferring to PBS. Ex vivo MRI of mouse brain specimens was performed on a horizontal 7 Tesla MR scanner (Bruker Biospin, Billerica, MA, USA) with a triple-axis gradient system. Images were acquired using a quadrature volume excitation coil (72 mm inner diameter) and a receive-only 4-channel phased array cryogenic coil. The specimens were imaged with the skull intact and placed in a syringe filled with Fomblin (perfluorinated polyether, Solvay Specialty Polymers USA, LLC, Alpharetta, GA, USA) to prevent tissue dehydration (33). Three-dimensional diffusion MRI data were acquired using a modified 3D diffusion-weighted gradient- and spin-echo (DW-GRASE) sequence (34) with the following parameters: echo time (TE)/repetition time (TR) = 30/400ms; two signal averages; field of view (FOV) = 12.8 mm x 10 mm x 18 mm, resolution = 0.1 mm x 0.1 mm x 0.1 mm; two non-diffusion weighted images (b0s); 30 diffusion encoding directions; and b = 2,000 and 5,000 s/mm2, total 60 diffusion weighted images (DWIs). Co-registered T2-weighted and magnetization transfer (MT) MRI data were acquired using a rapid acquisition with relaxation enhancement (RARE) sequence with the same FOV, resolution, and signal averages as the diffusion MRI acquisition and the following parameters: T2: TE/TR=50/3000 ms; MT: TE/TR=8/800ms, one baseline image (M0) and one MT-weighted (Mt) images with offset frequency/power = -3 KHz/20 uT were acquired. The total imaging time was approximately 12 hours for each specimen. For the sas4-/-p53-/- and littermate controls (n=4/4, P28), PGSE and OGSE diffusion MRI data were acquired with the protocol described in (17) and a spatial resolution of 0.1 mm x 0.1 mm x 0.1 mm. All 3D        MRI data were interpolated to a numerical resolution of 0.06 mm x 0.06 mm x 0.06 mm to match the resolution of our MRI-based atlas. 

Magnetization transfer ratio (MTR) images were generated as MTR=(M0-Mt)/M0. From the diffusion MRI data, diffusion tensors were calculated using the log-linear fitting method implemented in MRtrix (http://www.mrtrix.org) at each voxel, and maps of mean and radial diffusivities and fractional anisotropy were generated,  The mouse brain images were spatial normalized to an ex vivo MRI template (35) using the Large Deformation Diffeomorphic Metric Mapping (LDDMM) method implemented in the DiffeoMap software (www.mristudio.org). The template images had been normalized to the Allen reference atlas using landmark-based image mapping and LDDMM.

Histological data. From the Allen moue brain atlas, single subject 3D Nissl data and 3D AF data (n=100), which were already registered to the ARA space, were down sampled to 0.06 mm isotropic resolution. 

Registration of MRI and histological data: Group average 3D MRI data in our previously published mouse brain atlas(35) were first spatially normalized to the ARA space. Briefly, fourteen major brain structures (e.g., cortex, hippocampus, striatum) in the atlas MRI data were manually segmented following the structural delineations in the ARA. Voxels that belong to these structures in the MRI and average 3D AF data in the ARA (down sampled to 0.06 mm isotropic resolution) were assigned distinct intensity values, and a diffeomorphic mapping between the discretized atlas MRI and ARA AF data was computed using LDDMM. The mapping was then applied to the original atlas MRI data to generate an MRI template registered to the ARA space. Using dual-channel LDDMM based on tissue contrasts in the average DWI and FA images and the MRI template, the 3D MRI data acquired in this study were accurately normalized to the ARA space. 

    NF and MBP stained images of the C57BL/6 mouse brain were downloaded from the ARA reference dataset. Images with major artifacts or tissue loss were excluded. Small tissue tearing and staining artifacts were removed using the Inpainting feature implemented in the photoshop heading brush tool (www.adobe.com), and dark voxels in the ventricles were replaced by the average intensity values of the cortex to match MRI data. The repaired images were down-sampled to an in-plane resolution of 0.06 mm/voxel. For each 2D histological image, the best-matching MRI section in the MRI template was identified, and a coarse-to-fine alignment from histology to MRI using affine transform and landmark-based image warping tool in ImageJ (https://imageJ.nete/BUnwarpJ). The aligned 2D sections were then assembled into a 3D volume and mapped to the MRI template using LDDMM (between NF/MBP and FA) to further improve the quality of registration.