Although structural MRI is limited to image resolution in the millimeter range and cannot directly measure individual axon diameters, its sensitivity to microstructural geometries goes beyond the image resolution. Indeed, MRI-based axon diameter methods have been presented over the last decade, but they still carry an inherent bias that limits their clinical application. We are working on a new type of axon diameter models that overcome these known biases to achieve a higher precision. However, we experienced that the classical approach to constructing biophysical models based on our anatmical assumptions becomes unreliable and unstable for this new type of axon model. The reason was related to the fact that our new axon model aimed at extracting more detailed anatomical information (i.e. more parameters) from the MRI signal. If the anatomical details had too similar MRI signal fingerprints, the model failed.
Instead, we applied machine learning (ML)-based models that are not limited by the assumptions in classical biophysical models, demonstrating outstanding performance—though their performance degrades dramatically when applied to real MRI data. Currently, we are testing a new type of data preprocessing step to overcome this degraded performance observed in the ML models, and the results are promising.
The next question is: how do we know what the next generation of axon models should look like when we cannot directly visualize axons and myelin? Specifically, this is an important question when training and testing ML-based models. We obtain this geometrical information using X-ray Nano Holotomography (XNH), a 3D imaging technique involving experimental setups such as beamlines installed at large-scale synchrotron facilities like the ID16A at ESRF in Grenoble, France, or the P10 at DESY in Hamburg, Germany. The imaging datasets are massive in size, and we have had to develop new image analysis strategies to handle and segment the collected XNH data into tissue classes containing axons, cells, blood vessels, and myelin, for subsequent quantification of the shape, size, and organization of these anatomical structures.
Using previously collected XNH image data of a monkey brain with 75-nanometer resolution, we have established such an image processing pipeline. This includes deep learning-based segmentation methods that, from a very sparse manually segmented 3D training dataset and without specific preprocessing of the data, can generalize the segmentation of entire XNH 3D image volumes into classes of axons, myelin, blood vessels, and cells. Now knowing how the real microstructure environment looks in 3D, we have developed a software tool named the White Matter Generator (WMG) (doi:10.3389/fninf.2024.1354708) to synthetically generate 3D meshes of realistic microstructure environments for diffusion MRI simulations. As illustrated in Figure 1, the WMG software can be used to generate various type of realistic 3D microstructural environments of both normal and diseased cases, such as demyelination or axonal degeneration, for training, testing, and validation of both ML and biophysical microstructure models.
The myelin wrapped around the axon is a key component that ensures the boosted conduction velocity of signal transmission along the axons. Today, MRI-based methods can provide insight into the amount of myelin within an MRI voxel. When combined with other MRI-based techniques, such as axon diameter methods, these can also provide insight into the so-called g-ratio—a single metric describing the ratio between axon diameter and myelin thickness. However, for these metrics it is challenging to differentiate whether a disease is degenerating the myelin or if the brain is in a re-establishment phase, repairing degenerated myelin after a disease attack—valuable information for doctors to better understand the state of such diseases.
We have explored a new myelin-based MRI method that potentially enables such insights into disease processes. Tests applied to animal models include induced focal demyelination, followed over time by remyelination. When compared with histological data from the same brain, these results show promising potential.
