P1.48. CHARACTERIZATION OF THE CEREBRAL CORTEX IN THE NON-HUMAN PRIMATE BRAIN:
A DEEP LEARNING MODEL INTERPRETABILITY APPROACH
Adam Datta1, Agata Kulesza1, Marcin Syc1, Marcello G. P. Rosa2, Piotr Majka1,2
1 Institute of Experimental Biology, Laboratory of Neuroinformatics, 3 Pasteur St., Warsaw, Poland
2 Biomedicine Discovery Institute, Monash University, Department of Physiology and Neuroscience Program, Wellington Rd, Clayton VIC 3800, Australia
INTRODUCTION: Understanding how the cerebral cortex processes information involves laborious and knowledgeable characterization of its cytoarchitectonic properties. While it has been investigated for over a century, there is still no consensus regarding its structural and functional parcellation. Ongoing development of deep learning techniques provides promising support in the rapid processing of large amounts of high-resolution microscopic images, bringing a chance to alleviate these obstacles.
AIM(S): The study has two primary goals: to expand knowledge about the common marmoset (Callithrix jacchus) cerebral cortex features through quantitative characterization based on observer-independent segmentation, and to identify delineation criteria established by the U-Net deep learning model while assessing its biological plausibility against expert neuroanatomical knowledge.
METHOD(S): To ensure a valid delineation of the cortex into layers and areas, we: (1) estimated neuronal density and size, (2) extracted one-dimensional cortical profiles, (3) trained a deep-learning model to segment and classify cortical profiles, (4) applied Gradient-weighted Class Activation Mapping (Grad-CAM) to investigate model's decisions, and (5) measured cortical depth along 3D profiles to quantify laminar thickness and spatial organization across the cerebral cortex.
RESULTS: The model applied to a dataset of Nissl-stained coronal sections of the marmoset brain was able to recognize layers and assist in manually performed delineation. Further evaluation revealed increased performance when neural density and size estimates contributed to the training process. Additionally, the model was capable of correctly assigning area labels to each profile based on learned features.
CONCLUSIONS: Beyond automating identification and quantification of individual layers and areas, our solution provides valuable insight into cytoarchitectonic properties by highlighting the most contributive parts of the cerebral cortex.
FINANCIAL SUPPORT: NCN SONATA 2019/35/D/NZ4/03031