Virtual fluorescence labeling of mitochondria in phase images using a network with unpaired datasets

Abstract

Mitochondrial dynamics and morphology are closely linked to many cellular processes and pathologies. Conventional fluorescence microscopy allows for imaging of mitochondria selectively using fluorescent labeling, and consequently it suffers from phototoxicity and limited fluorescence channels. Quantitative phase contrast microscopy (QPCM) allows for imaging of tens of organelles in a label-free manner, yet it lacks the ability to distinguish specific organelles. In this work, we introduce an unsupervised deep learning model, entitled phase to fluorescence generative adversarial network (P2F-GAN). This model allows for virtual fluorescence labeling of mitochondria in QPCM images, eliminating the need for time-consuming paired training data acquisition. Utilizing an attention module and a customized loss function, P2F-GAN allows for virtual labeling of mitochondria (demonstrated as an example), achieving a structural similarity of 0.88, a Pearson correlation coefficient of 0.86, and a Dice coefficient of 0.84. The capability of the method has been demonstrated for tracking mitochondria in both physiological conditions and pharmacological interventions. The proposed method can be extended to other subcellular structures and invites many applications.