NEURODRIFT : 3D DEEP LEARNING FOR NEURAL SHAPE ANALYSIS IN NEUROSCIENCE

Abstract

3D Neural structures extracted through volume electron microscopy (EM) require careful analysis and labeling to understand their phenotype which store functional specialty and pathology. Identifying nanometric-scale 3D reconstructions of cellular and subcellular cells in neural tissue is a challenging task due to the high variability of structures and limited availability of labeled data, making scalable and data-efficient computational approaches necessary. This thesis investigates geometric deep learning methods, with a supervised and self-supervised approach, for morphology-driven analysis of ultrastructural EM data. Instead of relying on stagnant handcrafted geometric descriptors, state of the art (SOTA) deep learning methods, DiffusionNet and Laplacian2Mesh, are trained using the labeled neural objects within each class and tested on unseen labeled data to measure prediction accuracies. To address the limitation of scarce labels, a self-supervised learning approach, based on MoCo, is further explored, where the model is pretrained on unlabeled mesh data using contrastive learning with geometry-preserving augmentations, evaluated and finetuned with frozen KNN and linear probing. The results show that deep learning models effectively capture the intrinsic structural properties of neural morphologies. A key finding is that self-supervised pretraining generates geometry-aware embeddings capable of matching fully supervised baseline accuracy on downstream classification tasks. NeuroDrift demonstrates that combining geometric deep learning with self-supervised learning provides a scalable and data-efficient framework for analyzing ultrastructural neural morphology. The findings support the use of morphology-driven approaches alongside connectivity analysis and enable applications such as classification, clustering, and phenotype discovery in neuroscience.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering