Combining Multiple Annotations to Count Cells in 3D Cardiovascular Immunofluorescent Images

摘要

Cell quantification in 3D atherosclerotic plaque immunofluorescent images is critical in improving understanding of plaque composition and possibly preventing rupture. Manual annotation of these cells is required to build predictive models that classify cell types within the lesion. Due to the difficulty and the amount of time required to label a substantial number of cells, researchers must often aggregate multiple annotation strategies with various limitations. We present several methods to create segmentation masks to include polygon outlines, basic shapes, and small dots. The combination of these methods leverages the nuclei polygon annotations that must be linked to the other cellular markers. A U-Net segmentation model was trained with these annotations and used for evaluation. The models achieved a test set total cell count percent error of -2.7 when assessing total cell count alone; and 5.3, and 1.4 when counting ACTA2+ or LGALS3+ cells respectively. Further analysis of these cell counts with patient phenotype data may lead to findings that ultimately reduce heart attack and stroke rates.