A practical extraction and spatial statistical pipeline for large 3D bioimages

ABSTRACT/SUMMARYDespite much being known about the molecular regulation of hematopoiesis, little is understood about how hematopoietic cells are organized within bone marrow (BM) tissue. Recent advances in microscopy have led to the creation of increasingly detailed images of murine hematopoietic tissue. Accurate, efficient, and informative methodologies to extract and analyze the large amount of data generated are, however, still lacking. Indeed, cells are very densely packed in the bone marrow and therefore difficult to efficiently and accurately segment. In addition, currently employed statical analyses of cellular localization are generally unsuitable for the investigation of interactions between more than two cell types. To overcome these limitations, we developed PACESS, a readily applicable method based on neural network classification of hundreds of thousands of cells in thick 3D bone marrow samples, and a combination of statistical techniques to assess the spatial interactions between multiple cell types. To validate this approach, we used it to investigate the spatial organization of T cells, megakaryocytes and leukemic cells. We demonstrate that the presence of large clusters of leukemic cells affects the distribution of both T cells and megakaryocytes, albeit differently, resulting in the generation of previously unrecognized, unique microenvironments adjacent to each other within the same bone marrow cavity. We believe that this approach can contribute to unravel the BM cellular organization. MOTIVATIONThe organization of diverse hematopoietic cells within bone marrow tissue remains unclear. Recently developed tissue clearing methods enable the generation of large, 3D, single cell resolution microscopy images datasets, but the bottleneck in their analysis lies in both the identification and classification of cells, and in statistical analyses to probe their spatial relationships. We present a workflow (PACESS) that takes advantage of a convoluted neural network to identify and classify cells in 2D coupled with an automated method that extrapolates to 3D, followed by a combination of spatial statistics to classify tissue regions based on each cell types density, and logistic regression to test whether the relative abundance of cell types may be explained by reciprocal dependencies. Finally, we provide a combined measurement of the abundance of all cell types in a 3D map, highlighting regional variations found within the tissue. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=195 SRC="FIGDIR/small/521787v3_ufig1.gif" ALT="Figure 1"> View larger version (51K): org.highwire.dtl.DTLVardef@4bcae8org.highwire.dtl.DTLVardef@1f676bcorg.highwire.dtl.DTLVardef@1ed38d3org.highwire.dtl.DTLVardef@12de53b_HPS_FORMAT_FIGEXP M_FIG C_FIG