Computational Methods for Delineating Multiple Nuclear Phenotypes from Different Imaging Modalities
StatisticsView Usage Statistics
Characterizing histopathology or organoid models of breast cancer can provide fundamental knowledge that will lead to a better understanding of tumors, response to therapeutic agents, and discovery of new targeted therapies. To this aim, the delineation of nuclei is significantly interesting since it provides rich information about the aberrant microanatomy or colony formation. For example, (i) cancer cells tend to be larger and, if coupled with high chromatin content, may indicate aneuploidy; (ii) cellular density can be the result of rapid proliferation; (iii) nuclear micro-texture can be a surrogate for fluctuation of heterochromatin patterns, where epigenetic aberrations in cancers are sometimes correlated with alterations in heterochromatin distribution; and (iv) normalized colony formation of cancer cells, in 3D culture, can serve as a surrogate metric for tumor suppression. These evidences suggest that nuclear segmentation and profiling is a major step for subsequent bioinformatics analysis. However, there are two barriers which include technical variations during the sample preparation step and biological heterogeneity since no two patients/samples are alike. As a result of these complexities, extension of deep learning methodologies will have a significant impact on the robust characterization and profiling of pathology sections or organoid models. In this presentation, we demonstrate that integration of regional and contextual representations, within the framework of a deep encoder-decoder architecture, contribute to robust delineation of various nuclear phenotypes from both bright field and confocal microscopy. The deep encoder-decoder architecture can infer perceptual boundaries that are necessary to decompose clumps of nuclei. The method has been validated on pathology section and organoid models of human mammary epithelial cells.