Deep-learning applications in the automated image analysis of giant unilamellar vesicles

Recent progress in the field of machine learning allowed scientists to develop software to analyze a great amount of image data efficiently. Trained artificial neural networks can segment, analyze, and classify the target objects in complex images. Face recognition technology in our cell phones is an everyday experience now. We have been developing automated image analysis software to analyze lipid vesicle images efficiently. Giant unilamellar vesicles (GUV), characterized by a diameter greater than 1 µm are routinely used for reconstitution studies. They are used to study a wide range of physiological activities involving lipid membranes. One of the great advantages of the GUV system is that they can be readily imaged, and studied by optical microscopy due to their size much greater than the resolution of visible light's diffraction limit. Fluorescence microscopy image analysis can reveal the protein-lipid interaction, phase separation on the membranes, and mechanical deformation of the membranes. Such analysis typically involves the manual analysis by researchers to identify the target vesicles to analyze and perform quantitative analysis. Our approach to introducing deep-learning can automate the process to help researchers to analyze a great amount of vesicle image data efficiently. Our work shows that with some training, scientists can introduce the deep-learning as a module for their image analysis for various purposes. We discuss the progress in the field of software analysis of vesicle images and show our example works of deep learning applications. Deep-learning assisted identification, and classification of vesicle states and their connection with quantitative analysis will be shown. We suggest practical ways to implement it for various purposes of image analysis. We envision such modules can be potentially introduced in the classroom to teach future scientists to learn machine learning applications in biophysics.