Autoencoders for unsupervised analysis of rat myeloarchitecture.

Abstract

Quantitative assessment of brain histology is often constrained by predefined feature sets and labor-intensive manual annotations. To overcome these limitations, we employed unsupervised deep learning to automatically extract and quantify tissue organizational patterns from myelin-stained rat brain sections without the need for prior labeling. We evaluated nonlinear convolutional autoencoders (AEs) against linear principal component analysis (PCA) and DINOv3, a self-supervised foundation model applied without task-specific fine-tuning, for feature representation, followed by clustering with Gaussian mixture models. Compared to both PCA and DINOv3, AEs better preserved fine axonal architecture and produced more consistent and interpretable tissue clusters across hierarchical levels. The resulting clusters revealed anatomically meaningful tissue organization, including different white matter densities and grey matter subregions. When applied to tissue from sham and mild traumatic brain injury animals, AE-derived features also captured pathology-related alterations, such as white matter loss and injury-specific microstructural changes. These findings suggest that task-specific unsupervised training on domain-specific histological data can automatically characterize tissue organization at multiple scales and detect pathological changes, offering a scalable, annotation-free approach to computational neuropathology.