In silico prediction of cellular organelles from computationally super-resolved (SR) phase-modulated optical micrographs

Differential interference contrast (DIC) & phase contrast microscopy (PCM) represent 2 widely-assimilated optical microscopical imaging modalities often employed for live cell imaging (LCI) applications. Although both of these approaches have marked advantages over traditional brightfield (BF) microscopy (they dont require staining and provide real time visualization of cellular dynamics & metabolomics), there is a need to often supplement these modalities with epifluorescent microscopical imaging for the detection of specific organelles and structures in vivo (a process known as fluorescence combination/colocalization microscopy). Nonetheless, epifluorescent microscopy (both widefield and confocal) rely on the use of fluorophores (specialized fluorescent molecules and/or fluorescent proteins) for labelling key structures and features in the cell, which are prone to several issues, including photobleaching or cross-talk (amongst others). In this context, we seek to develop a novel deep neural network (DNN)-based approach aimed at predicting the location of 3 specific organelles (the cell nucleus, mitochondria & Golgi Apparatus) from acquired PCM & DIC images which have been super-resolved (SR) using our previously developed (& published) O-Net & {Theta}-Net model architectures. The model-generated images depict relatively close correlation with the ground truth images, implying that the O-Net & {Theta}-Net model architectures serve as viable frameworks for DNN model sculpting and training for both the purposes of image SR in optical microscopy as well as feature localization in these SR micrographs. We thus surmise on the potentiality of our proposed O-Net & {Theta}-Net DNN architectures for the development of models to be deployed in fully-automated image analysis pipelines in biomedical and healthcare diagnostics in the near future.