EXPLAINABLE AI-BASED CHARACTERIZATION OF LAMINAR ORGANIZATION IN THE COMMON MARMOSET CEREBRAL CORTEX

The structural organization of the cerebral cortex underpins its functional properties. However, it is not fully understood how cortical lamination affects information processing. In recent years, deep-learning methods facilitated cortical segmentation and classification into areas. Nevertheless, it is not a common practice to apply these kinds of methods beyond areas with well-described structure and function. To alleviate these challenges, we propose a semi-automated method for segmentation of the marmoset monkey cerebral cortex into areas and layers. The method integrates expert-based delineations with explainable deep-learning methods. Our workflow comprises (1) extraction of one-dimensional cortical profiles consisting of image intensity, estimates of neural density and size, (2) training of a U-Net convolutional neural network to segment out and classify cortical layers and areas, (3) Gradient-weighted Class Activation Mapping (Grad-CAM) to explore model's decisions, (4) and 3-D reconstruction to investigate cytoarchitectural properties of selected neocortical areas. The results demonstrated increased model performance when all the aforementioned cytoarchitectonic features were included in the training process. The model scored Jaccard coefficient of approximately 95% in the validation and 86.5% in the holdout dataset. However, the performance differed between individual layers. We observed that the set of features influenced mostly the segmentation accuracy of layers 2 and 4. Moreover, Grad-CAM highlighted contributory regions in profiles associated with the predicted area class. In summary, our method not only allows for semi-automated segmentation of cortical layers and classification into areas, but also introduces an innovative way to explain model’s decisions at different levels of laminar organization.