SELF-SUPERVISED LEARNING FOR AUTOMATED TRACKING OF NEUROANATOMICAL STRUCTURES

Quantitative analysis of neuroanatomical structures, such as axons, dendrites or dendritic spines, is essential for understanding structural plasticity in the brain. However, manual identification and tracking of these structures in microscopic images is time-consuming and subject to substantial inter-rater variability, limiting reproducibility and scalability of longitudinal studies. Modern supervised deep learning-based approaches typically depend on large amounts of manually annotated training data, which are costly to obtain and may themselves contain systematic errors.

We present a self-supervised deep learning framework for the reidentification of neuroanatomical structures in three-dimensional microscopic image stacks. Using two-photon microscopy data, we can show that our framework can create location-specific embedding representations without requiring manually labeled training data. Correspondence across stacks of the same scene is established by comparing these embeddings, enabling identification of the same structure across imaging sessions.

Preliminary results indicate robust matching performance between individual sessions, with a 95% confidence interval of 0.7 µm between predicted and ground-truth locations. These results demonstrate that self-supervised representation learning can support accurate, reproducible, and annotation-free tracking of fine neuronal structures. Our approach provides the basis for longitudinal studies of structural plasticity and spine dynamics, facilitating quantitative analysis of neuronal morphology across time without reliance on manual labeling.