TL;DR
This paper introduces SpIC3D, a novel 3D imaging technique that achieves unprecedented resolution of the spinal cord and its nerve roots, enabling detailed anatomical characterization and computational modeling. This advancement is crucial for understanding spinal cord function, diagnosing pathologies, and developing personalized treatments like spinal cord stimulation.
What Problem Does This Paper Address?
Visualizing the intricate neuroanatomy of the spinal cord, particularly the nerve rootlets and their complex organization at lower spinal levels, has been a significant challenge for traditional MRI techniques. This limitation hinders our ability to fully understand spinal cord function, diagnose diseases, and develop effective neuromodulation therapies. Existing methods often require extensive sample preparation, leading to tissue distortion, or lack the resolution needed to resolve fine structures.
Methods
The SpIC3D pipeline combines several advanced imaging and preparation techniques:
- Tissue Preparation: Samples are treated with a Gadolinium contrast agent to reduce T1 relaxation times, allowing for faster imaging. They are then immersed in Fomblin oil for magnetic susceptibility matching and to prevent dehydration.
- Multi-modal MRI Acquisition: SpIC3D utilizes a suite of MRI sequences, including:
- T2-weighted RARE: For general tissue contrast and visualization of white and gray matter.
- Fat-selective RARE: To enhance visualization of epidural fat.
- 19F MRI: A novel application using Fomblin to image the CSF void space.
- High-resolution Diffusion Tensor Imaging (DTI): Optimized with 19 diffusion directions to reconstruct and differentiate spinal roots and rootlets.
- Micro-CT: Used to capture the detailed structure of the vertebrae, which are kept intact throughout the process.
- Automatic Segmentation: A modified HarDNet neural network is trained to automatically segment six spinal tissues (gray matter, white matter, dorsal roots/rootlets, ventral roots/rootlets, CSF, and epidural fat) from the acquired images.
- Computational Modeling: The high-resolution 3D volumes and fiber trajectories derived from SpIC3D are used to build realistic computational models of the spinal cord for simulating neural responses to electrical stimulation.
Datasets
The study utilized postmortem spinal cord samples from:
- Cats: Six perfused and formalin-preserved cat spines (C01-C06).
- Humans: One cadaveric human thoracic-lumbar spine sample (H01).
These samples were used for both imaging acquisition and for training the automatic segmentation neural network.
Key Insights and Ideas
- Visualizing the Unseen: SpIC3D achieves ultra-high resolution, revealing fine details of spinal cord structures, including individual rootlets, that were previously difficult or impossible to visualize.
- CSF Imaging Breakthrough: The novel application of 19F MRI with Fomblin provides a way to image the CSF void space, improving contrast and reducing noise in other MRI sequences.
- Preserving Anatomy: Keeping the vertebral bone intact during imaging is critical for maintaining the natural anatomical organization of the spinal cord and its roots, a significant advantage over methods that require removing the cord from the vertebral column.
- Bridging the Gap: The technique combines the high resolution of postmortem imaging with the intact structure of in-vivo imaging, offering a powerful tool for creating detailed spinal atlases.
- Computational Model Realism: The high-fidelity anatomical data from SpIC3D enables the creation of significantly more realistic computational models of spinal cord stimulation, leading to more accurate predictions of neural activation and improved therapeutic outcomes.
Key Results
- Resolution: SpIC3D achieves resolutions as fine as 50 µm in-plane and 50-100 µm slice thickness.
- Quantification: The method allows for quantification of neuronal cell density in dorsal root ganglia, multi-segment identification of roots and rootlets, and volumetric reconstruction of spinal structures.
- Segmentation Accuracy: The automatic segmentation algorithm achieved human-level performance (Jaccard Index of 0.78, Hausdorff Distance of 3.65) after training on a relatively small number of slices.
- Computational Model Performance: A SpIC3D-derived computational model showed significantly higher accuracy in simulating electromagnetic fields and neuronal recruitment compared to a canonical model, with up to a 96.39% difference in E-field strength in critical areas like dorsal rootlets.
- Generalizability: SpIC3D was successfully applied to different spinal levels (cervical, thoracic, lumbar, sacral) in cats and also generalized to human samples.
Limitations and Future Work
- Postmortem Imaging: The current SpIC3D method is applied to fixed postmortem tissues. Future work aims to translate this to in-vivo imaging.
- Computational Cost: While the segmentation is fast, the overall imaging acquisition and processing pipeline can be time-consuming.
- Generalization to In-Vivo: The authors believe that the optimized gradient directions and b-values used in their DTI protocol could generalize to in-vivo tissues, paving the way for in-vivo applications.
- Personalized Models: The SpIC3D-derived models are intended to facilitate the development of personalized spinal cord stimulation protocols for improved safety and efficacy.
Summary updated