Speaker
Description
In recent years, three-dimensional polarized light imaging (3D-PLI) has opened new avenues for measuring and analyzing nerve fiber orientations in postmortem brains at the micrometer scale. The raw data consists of 18 or 9-channel images, while common derived formats include transmittance, retardation, inclination and computed fiber orientation maps. These measurements are highly detailed and complex, making interpretation and analysis of 3D-PLI data challenging. Thus, it would be beneficial to find local 3D-PLI texture features with reduced complexity that allow for grouping of regions with homogeneous fiber properties and provide good interpretability.
In this work, we build on the SimCLR framework for automatic generation of such highly descriptive 3D-PLI texture features, using a ResNet-50 as backbone. Training and test data were obtained from coronal whole brain sections of a 2.4-year-old, male vervet monkey brain. To demonstrate the descriptive power of the generated texture features, we implemented a border detection procedure to detect texture changes along the cortex around the calcarine and intraparietal sulci.
Several detected borders were confirmed by an expert to identify neuroanatomical plausible borders between brain areas. Therefore, we conclude that the self-learned texture features provide highly descriptive information content, while reducing the complexity of the 3D-PLI signal. Motivated by these findings, we aim to address the interpretation and analysis of the machine-learned features for 3D-PLI-based brain mapping in more depth.
