Validation of Structure Tensor Analysis for Orientation Estimation in Brain Tissue Microscopy

Abstract

Accurate localization of white matter pathways using diffusion MRI is critical to investigating brain connectivity, but the accuracy of current methods is not thoroughly understood. A fruitful approach to validating accuracy is to consider microscopy data that have been co-registered with MRI of post mortem samples. In this setting, structure tensor analysis is a standard approach to computing local orientations for validation. However, structure tensor analysis itself has not been well-validated and is subject to uncertainty in its angular resolution, and selectivity to specific spatial scales. In this work, we conducted a simulation study to investigate the accuracy of using structure tensors to estimate the orientations of fibers arranged in configurations with and without crossings. We examined a range of simulated conditions, with a focus on investigating the methods behavior on images with anisotropic resolution, which is particularly common in microscopy data acquisition. We also analyzed 2D and 3D optical microscopy data. Our results show that parameter choice in structure tensor analysis has relatively little effect on accuracy for estimating single orientations, although accuracy decreases with anisotropy. On the other hand, when estimating the orientations of crossing fibers, the choice of parameters becomes critical, and poor choices result in orientation estimates that are essentially random. This work provides a set of recommendations for researchers seeking to apply structure tensor analysis effectively in the study of axonal orientations in brain imaging data and quantifies the methods limitations, particularly in the case of anisotropic data. HighlightsO_LIStructure tensor estimates of homogeneous orientations are robust to parameter choice C_LIO_LIAccuracy in crossing-fiber regions is highly sensitive to parameter choice C_LIO_LIImage anisotropy has a large effect on accuracy in 3D images C_LIO_LISimulations provide guidelines for choosing parameters based on image characteristics C_LI