AbstractSpatial localisation of proteins dictates cellular function. Hence, visualisation of precise protein distribution is essential to obtain in-depth mechanistic insights into protein roles during cellular homeostasis, dynamic cellular processes, and dysfunction during disease. Labelling and staining of cells with protein specific antibodies is therefore a central and widely used technique in cell biology. However, unspecific binding, or cytoplasmic signals originating from the antibodies, make the distinction of the fluorescence signal from cellular structures challenging. Here we report a new image restoration method for images of cellular structures, using dual-labelling and deep learning, without requiring clean ground truth data. We name this method label2label (L2L). In L2L, a convolutional neural network (CNN) is trained with noisy fluorescence image pairs of two non-identical labels that target the same protein of interest. We show that a trained network acts as a content filter of label-specific artefacts and cytosolic content in images of the actin cytoskeleton, focal adhesions and microtubules, while the contrast of structural signal, which correlates in the images of two labels, is enhanced. We use an established CNN that was previously applied for content-aware image restoration, and show that the implementation of a multi-scale structural similarity loss function increases the performance of the network as content filter for images of cellular structures.
Label2label: Using deep learning and dual-labelling to retrieve cellular structures in fluorescence images
