Morphology and Connectivity of cells following Focused Ultrasound-mediated Blood-Brain Barrier Opening and Adeno-associated Virus Delivery

Abstract

Accurate identification and characterization of cells in the central nervous system (CNS) are essential for understanding brain function, cellular dynamics, and responses to experimental interventions. This is particularly important in contexts such as gene delivery, injury modeling, and studies of neuroinflammation, where different cell types exhibit distinct morphological signatures and activation states. Traditional methods for cell identification, based on immunofluorescence staining combined with manual confocal image acquisition and analysis, are time-consuming, costly, and prone to subjectivity, limiting scalability and reproducibility. This thesis presents the development of an AI-driven pipeline designed for automated cell detection, classification, and morphological analysis in CNS confocal fluorescent images. The pipeline integrates two deep learning-based models: Cellpose, a generalist segmentation tool for generating accurate cytoplasmic masks, and Detectron2, an instance segmentation framework based on Mask R-CNN, for simultaneous cell-type identification, signal quantification and detailed morphological and interaction analysis. Custom Cellpose models were developed for three major CNS cell types, astrocytes, microglia, and oligodendrocyte precursor cells (OPCs), enabling precise mask generation across varying staining conditions. Detectron2 was trained on annotated confocal images to perform instance-level segmentation and classification of astrocytes and microglia. Beyond cell detection, the pipeline assesses key morphological features, including area, solidity, eccentricity, circularity, and aspect ratio, for each detected cell. These characteristics enable the classification of cells into biologically relevant activation states, normal, intermediate, or activated, reflecting functional changes in response to experimental conditions or neurological pathologies. Furthermore, the workflow includes an interaction analysis in order to quantify astrocyte–microglia spatial relationships within the brain tissue. This pipeline was applied not only to standardized datasets but also to experimental images from mice subjected to focused ultrasound (FUS)-mediated blood-brain barrier (BBB) opening. Despite significant differences in magnification, anatomical region, and image scale between the training and experimental datasets, the pipeline reliably detected and classified GFP-labeled cells and produced detailed morphological and interaction reports. Overall, this approach demonstrates the effectiveness of AI-based methods for scalable, objective, and cost-efficient CNS cell analysis. By reducing the need for expensive immunostaining and manual analysis, the thesis suggests this pipeline as a powerful tool for researchers studying CNS biology, cellular responses to interventions, and tissue remodeling. Future improvements, including extension to additional cell types, incorporation of 3D morphological analysis, and the development of user-friendly software, would further expand its applicability across neuroscience research and experimental medicine.