(59b) Analysis of Microglia Morphology across Different Neuroinflammatory Rat Models

Animal models are common pre-clinical methods for studying pathological processes of brain diseases. Since no animal model matches the exact human presentation of any specific disease, most diseases are represented by multiple animal models that capture different aspects of cellular response. To combine the insights of different animal models, we propose an image processing-based cell quantification approach that enables comparison of pathological processes across models. However, effective combinatorial experimental modeling is complex, so we focus only on microglia - the immune cells of the brain - for the first application of our image processing pipeline. Microglia are of particular interest for animal models of neurological disease since they are a common measure of neuroinflammation. Neuroinflammation plays pathological roles in the ongoing impact of various neurological diseases such as by acting as the resident macrophage and by maintaining neural circuits¹. We apply a microglia morphology quantification pipeline that analyzes individual cell features and common populational cell shapes to images acquired from three separate models of neuroinflammation. Two imaging data sets are obtained from cultured organotypic whole hemisphere brain slice models from rats: (1) an oxygen-glucose deprivation (OGD) model and (2) a lipopolysaccharide (LPS)-induced neuroinflammation model. Our third data set is obtained from the Vannucci model of HI in rats². By analyzing the microglia morphology of all three models using the same image processing pipeline, we can study the reactivity of microglia across models and compare specific benefits or features of each model. Our microglia morphology analysis pipeline quantifies microglia at the individual level through geometric features including perimeter, circularity, and area and at the population level through principal components analysis and k-means clustering that allows us to group microglia into the most common shapes. Our multi-level microglia morphology analysis across models of neuroinflammation provides a basis for a combinatorial model approach to study unique aspects of disease models. The combinatorial approach will improve pre-clinical disease characterization of pathological processes by increasing the number of features of microglia morphology that can be quantified and then exploring the extent to which those features change.

Methods:

Image Processing

We begin with immunofluorescent neural images of various brain regions across the three models. After image acquisition, the steps of the image processing pipeline are to (a) segment every cell from all images using a threshold-based method, (b) compute geometric features of every cell – i.e. perimeter, circularity, aspect ratio – using the Sci-Kit Image package in Python, and (c) apply the Visually Aided MorphoPhenotyping Image REcognition (VAMPIRE)³ package to apply principal components analysis and k-means clustering to determine the five most common shapes and (d) visualize the data as scatter plots of every cell for all geometric features and heat maps of population trends (Figure 1).

Figure 1. Microglia morphology analysis pipeline showing how cells are analyzed from multiple regional images. Cells go through two levels of analysis: (top row) individual: geometric feature analysis and (bottom row) population: group clustering and then are visualized as individual results and as population trends. The data is compared by either considering the whole slice as one group or the regions of each slice as different groups. After all slices have been quantified, inter-slice comparisons for every experimental group are computed.

Image Segmentation

Within each experimental group and each subset – i.e. image magnification, region, and sex differentiated groups – 1/3 of the data was randomly selected for manual segmentation. Manual segmentation and manual counting were carried out by three different researchers using ImageJ. After manual segmentation, ImageJ was also used to calculate the total cell area coverage of each image. Alongside manual segmentation, a set of thresholds – isodata, Li, mean, minimum, Otsu, triangle, Yen - were applied automatically using the Sci-Kit Image package in Python with the same set of data. After the thresholds were applied, cell count and total cell area were calculated using Python. The automated threshold-based segmentation scores were compared to the manual scores and the best threshold-based segmentation method was chosen for all data from each of the three HI models.

Data Splitting for VAMPIRE

For the train:test data split for the VAMPIRE package, we approached the data split in three ways: (1) all data together, (2) each model separately, and (3) cross model comparison. In all models, all data sets were bundled together and split into 80:20 train:test split and then all data sets were run using the compiled model. For each model separately, every disease model was used to create its own VAMPIRE model and then run on only that model’s data. For the cross-comparison model, we ran each disease model through the VAMPIRE shape modes created by the other two disease models to compare clustered shapes and reactivity in a multi-level way.

Data Visualization

Data was visualized in two overarching ways - within each model and across each model - and at two separate levels - individual features and population trends. Individual feature maps included scatter plots of geometric features of every single cell analyzed. Meanwhile, population trends included heatmaps of shape mode frequencies and heatmaps of changes both in respect to other models and in respect to the non-treated control.

Results:

Results from the two ex vivo models show that both models change microglial morphological features after treatment in regards to perimeter, area, circularity in comparison to the non-treated control groups. However, when comparing models, we see that the change shows regional differences and different extents of reactivity. Additionally, the models show relatively similar overall clustering shape modes with each model returning a shape that is clearly circular and other shapes that are different extents of branching as supported by previous studies in literature^{4, 5}. In population analysis, the models both show increased heterogeneity with treatment, but different extent of heterogeneity and shape mode reactivity in comparison to the non-treated control groups. Additionally, we completed the segmentation steps on different magnifications of the LPS model images that qualitatively show similar trends across magnification with increased insight to individual geometric features – especially branching - but decreased insight in population trends as magnification increases. Meanwhile, the Vannucci in vivo model shows decreased microglial count in the injured model with regional variations in microglia circularity⁶.

Conclusions:

Different etiologies of disease can affect cellular response in unique ways. For example, in response to neuroinflammation microglia contract branches and swell therefore increasing circularity. We can capture microglial geometric changes with our image processing pipeline. In all models of neuroinflammation, we observed different extent of changes in microglial shape features and different population reactivity when comparing the injured groups to both non-treated control groups and each other. Since changes in microglia morphology provide insight into neuroinflammatory processes, comparing microglia across models may be a way to explore combinatorial model approaches for better modeling of human disease processes. By characterizing unique aspects of microglia morphology both across brain regions and across models, we learn specific features of microglia that can be teased out with future experiments. For example, by quantifying extent of circularity change of microglial in injured groups and treatment, we establish a method for studying disease severity and treatment impact across models. By establishing methods of characterizing and comparing cellular response across different etiologies that result in similar pathological processes, we show promise in using combinatorial model approaches to better match human pathological processing.

References:

1. Muzio, L.; Viotti, A.; Martino, G., Microglia in Neuroinflammation and Neurodegeneration: From Understanding to Therapy. Front Neurosci 2021, 15, 742065.
2. Rumajogee, P.; Bregman, T.; Miller, S. P.; Yager, J. Y.; Fehlings, M. G., Rodent Hypoxia-Ischemia Models for Cerebral Palsy Research: A Systematic Review. Front Neurol 2016, 7, 57.
3. Phillip, J. M.; Han, K.-S.; Chen, W.-C.; Wirtz, D.; Wu, P.-H., A robust unsupervised machine-learning method to quantify the morphological heterogeneity of cells and nuclei. Nature Protocols 2021.
4. Nguyen, N. P.; Helmbrecht, H.; Ye, Z.; Adebayo, T.; Hashi, N.; Doan, M.-A.; Nance, E., Brain Tissue-Derived Extracellular Vesicle Mediated Therapy in the Neonatal Ischemic Brain. International Journal of Molecular Sciences 2022, 23 (2), 620.
5. Wood, T. R.; Hildahl, K.; Helmbrecht, H.; Corry, K. A.; Moralejo, D. H.; Kolnik, S. E.; Prater, K. E.; Juul, S. E.; Nance, E., A ferret brain slice model of oxygen–glucose deprivation captures regional responses to perinatal injury and treatment associated with specific microglial phenotypes. Bioengineering & Translational Medicine 2021.
6. Joseph, A.; Wood, T.; Chen, C.-C.; Corry, K.; Snyder, J. M.; Juul, S. E.; Parikh, P.; Nance, E., Curcumin-loaded polymeric nanoparticles for neuroprotection in neonatal rats with hypoxic-ischemic encephalopathy. Nano Research 2018, 11 (10), 5670®C5688.