Classification of Diffuse Subcellular Morphologies
Neelima Pulagam
Marcus Hill
Mojtaba Fazli
Rachel Mattson
Meekail Zain
Andrew Durden
Frederick D Quinn
S Chakra Chennubhotla
Shannon P Quinn
Characterizing dynamic sub-cellular morphologies in response to perturbation remains a challenging and important problem. Many organelles are anisotropic and difficult to segment, and few methods exist for quantifying the shape, size, and quantity of these organelles. The OrNet (Organelle Networks) framework models the diffuse organelle structures as social networks using graph theoretic and probabilistic approaches. Specifically, this architecture tracks the morphological changes in mitochondria because its structural changes offer insight into the adverse effects of pathogens on the host and aid the diagnosis and treatment of diseases; such as tuberculosis. The OrNet framework offers a segmentation pipeline to preprocess confocal imaging videos that display various mitochondrial morphologies into social network graphs. Earlier methods of anomaly detection in organelle structures include manual identification by researchers in the biology domain. Although those approaches were successful, manual classification is time consuming, tedious, and error-prone. Existing convolutional architectures do not have the capability to adapt to general graphs and fail to represent diffuse organelle morphologies due their amorphous characteristic. Thus, we propose the two different methods to perform classification on these organelles that captures their dynamic behaviors and identifies the fragmentation and fusion of mitochondria. One is a graph deep learning architecture, and the second is an approach that finds a graph representation for each social network and uses a traditional machine learning method for classification. Recent studies have demonstrated graph neural network models perform well on time-series imaging tasks, and the graph architectures are better able to represent amorphous and spatially diffuse structures such as mitochondria. Alternatively, much research has established traditional machine learning methods to be promising and robust models. Testing and comparing different architectures and models will effectively improve the robustness of categorizing distinct structural changes in subcellular organelle structures that is very useful for identifying infection patterns, offering a new way to understand cellular health and dynamic responses.
DOI
10.25080/majora-1b6fd038-00f