User-friendly deep learning-based morphometric unmixing of multiplex 3D imaging data

Immunofluorescence imaging techniques allow us to unravel the cellular morphology and microenvironmental components, which permits us to examine and characterize cell populations and understand the tissues’ architecture. However, the number of fluorescent markers used for visualization is constrained by the spectral emission of fluorophores that gives rise to spectral overlap, challenging the discrimination between numerous markers simultaneously. In previous work, it has been proposed the use of a conditional generative adversarial network (cGAN), pix2pix, for shared fluorescent signal unmixing between two different markers based on their subcellular distribution. Nevertheless, despite deep learning methods, such as cGANs, being powerful tools for microscopy, there is an entry barrier to their application in the biomedical field due to the lack of coding expertise in many microscopy users that might benefit from these tools. Here we present ZeroCostDL4Mic-VirtualMultiplexing, a platform that implements cGANs for signal unmixing using the cloud-based resources of Google Colab. This platform allows users with no coding experience to train and use their models to perform virtual multiplexing using the pix2pix framework. To generalize the applicability of using pix2pix models for virtual multiplexing, we have studied the performance of these models trained with different types of ground truth data for signal unmixing in imaging that contains two markers with spectral overlap, demonstrating its ability to perform virtual multiplexing. Our study provides a tool for accessing this technology and demonstrated a proof of principle of our approach to generalize the use of these networks