Deep Learning-Based Automatic Segmentation and Analysis of Mitochondrial Damage by Zika Virus and SARS-CoV-2

Viruses can induce various mitochondrial morphological changes, which are associated with the type of immune response. Therefore, characterization and analysis of mitochondrial ultrastructural changes could provide insights into the kind of immune response elicited, especially when compared to uninfected cells. However, this analysis is highly time-consuming and susceptible to observer bias. This work presents the development of a deep learning-based approach for the automatic identification, segmentation, and analysis of mitochondria from thin-section transmission electron microscopy images of cells infected with two SARS-CoV-2 variants or the Zika virus, utilizing a convolutional neural network with a U-Net architecture. A comparison between manual and automatic segmentations, along with morphological metrics, was performed, yielding an accuracy greater than 85% with no statistically significant differences between the manual and automatic metrics. This approach significantly reduces processing time and enables a prediction of the immune response to viral infections by allowing the detection of both intact and damaged mitochondria. Therefore, the proposed deep learning-based tool may represent a significant advancement in the study and understanding of cellular responses to emerging pathogens. Additionally, its applicability could be extended to the analysis of other organelles, thereby opening up new opportunities for automated studies in cell biology.