Abstract Text: PURPOSE The analysis of microglial morphology provides important data that are associated with immunoresponse and pathogenesis following TBI. Microglia are known for high plasticity and heterogeneity in their morphological phenotypes that usually require the analysis of 3D confocal image stacks to comprehensively reveal their morphology. Recent advancements in deep learning image analysis provide a great potential to quantify these highly variable cells using easy-to-acquire 2D low-power histological images. The purpose of this study is to understand how slice thickness of histology images may affect the accuracy of quantifying microglial morphology in brain tissue following TBI using deep learning. METHODS 9 female 12-week-old SD rats had mild TBI using a 2m height/450g weight drop. The other 5 rats without TBI were served as controls. After 2 weeks following TBI, the brain tissues were processed for histology, sectioned at 40μm thickness, and stained for microglia by Iba1 antibody. The 3D Iba1 images were acquired at three brain regions (cortex, corpus callosum, thalamus) using a laser scanning confocal microscope (Zeiss 710, Oberkochen, Germany) with a Plan-Apochromat objective (20x air, NA=0.8). The 3D microglial morphology was reconstructed from the full stacks of the confocal images and analyzed by neuroimaging experts for cell identification and categorization. A composite deep learning system, combining YOLOv5 and UNet, was utilized to unbiasedly quantify and categorize microglial morphology appearing on the 2D images, reconstructed at 10, 20 and 40μm thicknesses. Each microglia cell was classified into one of six categories (ramified, hypertrophic, bushy, ameboid, rod, and hyper-rod) that cover a morphological spectrum from surveillant to activated, in gray matter and white matter. The results from the 2D low-power histological images were then compared to the ‘gold standard’ confocal 3D stack-derived morphology results to determine the quality of detection and categorization of morphology by the deep learning model. RESULTS Thin slices do not provide sufficient features about the cell’s size, processes, and shapes, so thinner tissue provided less reliable identification of the cell’s morphological phenotypes. On the other hand, thicker tissue provided more accurate microglial features allowing the deep learning model to categorize the cell morphology by feature extraction and pattern recognition. Compared to 3D analysis, percent agreement of the 2D results using deep learning categorization increased by 9% from 10µm to 20µm, and by 6% from 20µm to 40µm, in reflecting the inflammatory effects of TBI. DISCUSSION / CONCLUSION A reliable quantification of microglia on 2D images can greatly increase the ease of analyzing a large number of cells and reveal the regional difference at a lower cost. 3D confocal images provide the most accuracy in quantifying microglial morphology. However, 3D images are resource-expensive and can only be acquired from a small field of view that are not suitable when investigating inflammation across the entire brain, such as in the secondary injury of diffuse TBI, where microglial activation may vary greatly across different brain regions. Low power (~20x) 2D images can be acquired by a conventional microscope and are better suited for high-throughput analysis of the entire brain. When analyzing microglia on the 2D images, it is essential to determine the accuracy of the automated quantification model in a wide variety of real-life experiments. Based on the results, a slice thickness threshold for detailed identification of microglial morphological features in 2D imaging can be determined (>20µm). Above this threshold, the cell’s features can be resolved by the deep learning model comparably to the ‘gold standard’ 3D analysis. The findings of this study also provide a direction for the future of deep learning models to extrapolate the images acquired from different slice thicknesses by transfer learning, which may still reliably reveal microglial morphology without information loss.
