ParticipantsThe MRI data was collected from 14 healthy adult volunteers (2 females and 12 males; Age: 23.21 ± 0.89 years; Height: 175.43 ± 8.22 cm; Weight: 71.14 ± 11.72 kg) between June 2023 and December 2023. Participants were publicly and randomly recruited from students in Fudan University. The inclusion criteria required participants to have no reported history of spinal cord injury (SCI), chronic back pain or stroke. Additional exclusion criteria were applied based on general guidelines for MRI safety and tolerance. All participants provided written informed consent, including an agreement for their data to be shared openly in an anonymous form. The experiment was approved by the ethics committee of Fudan University (approval number: FE23166I).Image acquisitionThe MRI scans were conducted at the Zhangjiang International Brain Imaging Center of Fudan University, using a 3T whole-body MRI system (MAGNETOM Prisma, Siemens Healthineers, Erlangen, Germany) equipped with a 20-channel head coil, an 18-channel body coil, and a spine coil. To locate different spinal structures, three MRI sequences were employed. (i) First, T2-weighted TSE (T2-TSE) was employed, with the following parameters: field of view = 240 mm  × 240 mm, voxel size = 0.63  × 0.63  × 3.00 mm3, TR/TE = 3500/104 ms, flip angle = 160°, readout bandwidth = 260 Hz per pixel, and the scan time = 3 minutes 22 seconds. To obtain images of the whole spine, T2-TSE sequences were applied 4 times targeting different segments. (ii) Second, double echo steady state sequence (DESS) was employed, with the following parameters were: field of view = 243 mm  × 243 mm, voxel size = 1.27 mm isotropic, TR/TE = 10.95/3.86 ms, flip angle = 25°, readout bandwidth = 325 Hz per pixel, and the scan time = 2 minutes 23 seconds. (iii) Third, 3D high-resolution constructive interference in steady state (CISS) was employed with the following parameters: field of view = 288 mm  × 288 mm, voxel size = 0.35  × 0.35  × 1.80 mm3 or 0.30  × 0.30  × 2.00 mm3 depending on the length of the participant’s intumescentia lumbalis, TR/TE = 9.80/4.46 ms, flip angle = 50°, readout bandwidth = 305 Hz per pixel, turbo factor = 19 and the scan time = 29 minutes 30 seconds. The entire imaging protocol lasted approximately 1 hour including the time for preparation and localization.Image preprocessingThe acquired MRI scans were converted from DICOM to Neuroinformatics Informatics Technology Initiative (NIfTI) format using dcm2niix (v1.0.20240202)20 (https://www.nitrc.org/plugins/mwiki/index.php/dcm2nii:MainPage) and then organized following the Brain Imaging Data Structure (BIDS) format21 using dcm2bids (v3.1.1)22. Example results are shown in Fig. 1.Fig. 1Representative MRI data acquired from a healthy adult participant illustrating the human lumbosacral spine from multiple dimensions and the following postprocessing pipeline. T2-TSE sequence images delineate the spinal cord contour, while DESS sequence images highlight ganglion localization. Additionally, MRI images from the CISS sequence distinctly depict the distribution of spinal nerve roots in the lumbosacral spine. The geometry information was obtained through manual annotation and was subsequently utilized to automatically construct a comprehensive human lumbosacral model, encompassing structures such as the dura, cerebrospinal fluid, and the nerve roots spanning from L1 to S2.Image postprocessingBased on the MRI images acquired, the lumbosacral spine model for each subject was constructed. The pipeline is illustrated in Fig. 1. Annotation of the MRI images was performed using 3D Slicer (v5.4.0), involving two steps. Firstly, DESS and T2-TSE sequence images were annotated to locate ganglions at each target segment, aiding in determining the trajectory of nerve roots after exiting the intervertebral foramina. Secondly, annotations were carried out on the CISS sequence images to delineate the trajectories of individual nerve roots at each target segment, as well as to define the boundaries of the spinal cord white matter and cerebrospinal fluid and the position of the dura mater. In CISS images, nerve roots appear as black dots in a white background. Annotators traced the positions of nerve roots for each target segment on successive slices of CISS sequence images. The segment that each nerve root corresponds to was determined by its distance from the center of the spinal cord in each slice (i.e., closer proximity to the spinal cord center in a given slice indicates a lower spinal cord segment).Two trained annotators participated in the annotation process, resolving uncertainties through consensus after discussions. The final annotation results were obtained by averaging the annotations of the two annotators. The annotation differences between the two annotators for different nerve roots are shown in Fig. 2. The figure indicates that the uncertainty in annotations gradually increases from L1 to S2. This trend is due to the increasing density of nerve roots within the field of view from L1 to S2. Although the annotation differences for the S1 and S2 nerve roots are relatively larger, they are still smaller compared to the width of the Medtronic 5-6-5 paddle lead (10 mm)23 commonly used in EES, which is an important application relying on spine MRI. Therefore, the annotation differences between two annotators on S1 and S2 nerve roots can still be considered small enough and will not hinder related applications in practice. Furthermore, the placement of these implanted EES electrodes during implantation is primarily guided by the location of the L1 nerve root, the annotations of which between two annotators are highly consistent. In summary, the observed annotation discrepancies highlight the inherent complexity of these physiological structures and can provide a valuable reference for evaluating the performance of machine learning algorithms developed using this dataset. The original data corresponding to Fig. 2 and the standard deviation of the annotation differences are provided along with the dataset.Fig. 2Mean differences of annotations from two annotators for each subject, focusing on the L1 to S2 spinal cord nerve roots. Each data point represents the average annotation difference across different imaging slices. The differences were specifically measured by calculating the three-dimensional distances between the coordinates of corresponding points in the two sets of annotations.Subsequently, based on these annotations, individual lumbosacral models (Fig. 3) for each subject were constructed using the open-source modeling software Blender (v4.0.2). Manual adjustment was applied to rectify the intersection issues caused by annotation bias. The Blender script used for modeling can be accessed via Github (https://github.com/Joshua-M-maker/SpineNerveModelGenerator).Fig. 3Visualization of human lumbosacral models based on MRI data. The right 14 models were from 14 healthy adult participants in this work. Comparatively, the left model was from a spinal cord injury subject in another research26 without open-sourced MR images, markers and models. Each model incorporates anatomical structures including the dura, cerebrospinal fluid, and nerve roots extending from L1 to S2. The direction of the coordinate axes is indicated. Spinal cord models were aligned by the highest planes of the dura structure to exhibit individual variability.
