Linked recognition of antigen by T cells and antigen presenting cells (APCs) through direct cell-to-cell contact underlies most adaptive immune responses that both protect from infection and drive tissue damage in autoimmunity. To objectively assess this linked recognition of cells, our research studies the morphology and interaction of individual immune cells relative to inflammatory immune response. We utilize multi-channel immunofluorescence imaging (cellular confocal microscopy) in presentations of lupus nephritis incorporating nuclear stains and cell membrane immunofluorescent stains with deep convolutional neural networks (DCNNs) to classify multiple T cell and dendritic cell types. Specific to this dissertation, we investigate and develop computerized single-cell segmentation and characterization in cellular confocal microscopy images of lupus nephritis with the goal of quantifying interaction of immune cells to characterize immune response. The feasibility of localizing, segmenting, and classifying individual cells in multi-channel confocal microscopy images is highly dependent on image quality. In certain applications with good image quality and well-separated cells, segmentation can be trivial and accomplished via thresholding, watershed, or a collection of other well-established and studied heuristics; however, at the limit of poor image quality, densely packed cells, and complex image features, these techniques fail. It is at this limit that deep learning approaches excel. Using deep learning region-based segmentation techniques, we enable analysis of challenging image data that previously required laborious hand segmentation, based on multichannel information difficult for the human visual system to parse. We analyze cellular images from confocal microscopy of (a) fresh frozen tissue in a mouse model of T cell activation by dendritic cells and (b) fresh frozen tissue in kidney biopsies patients with lupus nephritis (c) paraffin embedded kidney biopsies from lupus nephritis patients. Our goal is to segment cells in order to quantify the interaction of immune cells to characterize in situ adaptive immunity. The main challenge of this work is refining segmentation of dense cells in a tissue medium, which is also applicable to general segmentation challenges of histopathological images. In murine fresh frozen images we achieve a segmentation performance with intersection over union (IOU) of 0.85, sensitivity of 0.88, and specificity of 0.92 across cell types. For human fresh frozen images an IOU of 0.70, sensitivity of 0.72, and specificity of 0.86 is achieved across cell types. Using DCNN segmentation and classification nuclear segmentation output, an area under the receiver operating characteristic curve (AUROC) of 0.82 is achieved for discriminating cell types in murine fresh frozen data and an AUROC of 0.64 is achieved for discriminating cell types in human fresh frozen data. Overall, we show that by utilizing multi-channel confocal microscopy and our novel DCNN pipeline, we achieve performances similar to two photon excitation microscopy (TPEM) in discriminating stable cognate from non-cognate T cell–dendritic cell interactions in mice and apply it also to human tissue. These data indicate that a quantitative analysis of many static two-dimensional images can approximate much of the information obtained from time-lapse three-dimensional videos of the same phenomenon. Additionally, since our analysis is performed on single-plane images of fixed tissue, we could use it to study human disease and identify important in situ APCs.




Downloads Statistics

Download Full History