Deep Learning–Based Segmentation of Morphologically Distinct Rat Hippocampal Reactive Astrocytes After Trimethyltin Exposure

Abstract

As regulators of homeostasis, astrocytes undergo morphological changes after injury to limit the insult in central nervous system (CNS). Trimethyltin (TMT) is a known neurotoxicant that induces reactive astrogliosis in rat CNS. To evaluate the degree of reactive astrogliosis, the assessment relies on manual counting or semiquantitative scoring. We hypothesized that deep learning algorithm could be used to identify the grade of reactive astrogliosis in immunoperoxidase-stained sections in a quantitative manner. The astrocyte algorithm was created using a commercial supervised deep learning platform and the used training set consisted of 940 astrocytes manually annotated from hippocampus and cortex. Glial fibrillary acidic protein-labeled brain sections of rat TMT model were analyzed for astrocytes with the trained algorithm. Algorithm was able to count the number of individual cells, cell areas, and circumferences. The astrocyte algorithm identified astrocytes with varying sizes from immunostained sections with high confidence. Algorithm analysis data revealed a novel morphometric marker based on cell area and circumference. This marker correlated with the time-dependent progression of the neurotoxic profile of TMT. This study highlights the potential of using novel deep learning-based image analysis tools in neurotoxicity and pharmacology studies.