Abstract
Axon diameter is an important neuroanatomical characteristic of the nervous system that alters in the course of neurological disorders such as multiple sclerosis. Axon diameters vary, even within a fiber bundle, and are not normally distributed. An accurate distribution function is therefore beneficial, either to describe axon diameters that are obtained from a direct measurement technique (e.g., microscopy), or to infer them indirectly (e.g., using diffusion-weighted MRI). The gamma distribution is a common choice for this purpose because it resembles the distribution profile of measured axon diameters. In this study, we compared a wide range of parametric probability distribution functions against empirical data obtained from electron microscopy images. We observed that the gamma distribution fails to accurately describe the main characteristics of the axon diameter distribution, such as location and scale of the mode and the profile of distribution tails. We found that the generalized extreme value (GEV) distribution consistently fitted the measured distribution better than other functions. Additionally, several 2-parameter distributions, including inverse Gaussian, log-normal, log-logistic, and Birnbaum-Saunders, outperformed the gamma distribution.
Introduction
Axon diameter is a critical neuroanatomical feature determining the conduction velocity of electrical signals in the brain. It changes dynamically during development and is often compromised in disorders like multiple sclerosis and autism. Because axons within a single nerve bundle vary significantly in size, researchers rely on statistical probability distribution functions (PDFs) to model these populations.
For years, the Gamma distribution has been the standard choice due to its non-negative nature and right-skewed profile, which superficially matches histological observations. However, inaccuracies in this mathematical model can lead to significant errors in biophysical models used in diffusion MRI (dMRI) to estimate axon size non-invasively. This study aims to identify a more accurate mathematical "template" for axon diameter populations across different species.
Methods and Materials
The researchers evaluated 16 different parametric probability distribution functions. To rank these models, they employed the Akaike Information Criterion (AIC), which balances goodness of fit against model complexity (number of parameters). This ensures that a more complex model is only preferred if its increased accuracy is statistically significant.
The study utilized high-resolution electron microscopy (EM) data from three species:
- Mouse: A large dataset of over 20,000 manually measured axons from the genu, body, and splenium of the corpus callosum.
- Human & Macaque: Previously published EM data used for cross-species validation.
Results and Discussion
The analysis revealed that the widely used Gamma distribution consistently fails to accurately describe the "mode" (the most frequent axon size) and the "tails" (the rare, very large axons) of the biological data.
In contrast, the Generalized Extreme Value (GEV) distribution emerged as the superior model across all brain regions and species. Its success suggests that axon populations may consist of distinct subpopulations, which the flexible GEV shape can capture more effectively. Interestingly, while the GEV was the best performer, the standard Extreme Value distribution was the worst, highlighting the necessity of the "generalized" formulation.
For applications requiring simpler models, several 2-parameter alternatives—including Log-normal, Inverse Gaussian, and Log-logistic distributions—significantly outperformed the Gamma distribution without increasing mathematical complexity.
Quantitatively, in the mouse dataset, GEV achieved an AIC score of -6159, a massive improvement over the Gamma distribution's score of -3463. Visual inspection confirms that the GEV curve tracks the empirical histogram much more closely, particularly at the peak and the long right tail.
Conclusion
While the GEV distribution provides the best fit, its requirement for three parameters poses a challenge for diffusion MRI signal inversion, which is sensitive to noise. However, the discovery of superior 2-parameter alternatives like the Log-normal distribution offers an immediate path to improving biophysical models. Future work should focus on integrating these more accurate distributions into frameworks like AxCaliber to enhance the precision of non-invasive "virtual biopsies" of brain microstructure.
Summary updated