Label-free segmentation of mitochondria for simultaneous morphological and metabolic studies

The dynamics in the mitochondrial structure and function are closely related to cellular health. Traditional fluorescence imaging techniques for observing mitochondria are limited by phototoxicity, photobleaching and staining artifacts. In this study, we propose RedoxSegNet, an AI-enhanced imaging platform that enables label-free segmentation of mitochondria for concurrent morphological and functional analysis without the aid of labeling dyes. Our approach uses high-resolution two-photon excitation fluorescence microscopy, in conjunction with a custom-built conditional diffusion model, to reconstruct mitochondrial features from NAD(P)H autofluorescence images. Subsequent segmentation through post-processing algorithms demonstrates a task-specific performance error of less than 6% on average compared to the mitochondria-stained images. Our trained model effectively extracts mitochondrial features from label-free images, thereby facilitating mapping of mitochondria-specific optical redox ratios. We find that our analysis elucidates metabolic heterogeneity both within and between the organelles. Further validation under mitochondrial stress conditions induced by carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone (FCCP) confirms that RedoxSegNet can capture dynamic mitochondrial fragmentation and heterogeneous metabolic response. Overall, these findings establish our technique as a non-invasive, reliable tool for investigating mitochondrial morpho-functional dynamics in native cellular environments.