Swin Transformer Based Neural Network for Protein Subcellular Localization Prediction from Quantitative Label-Free Imaging with Phase and Polarization (QLIPP) in Unlabeled Live Cells and Tissue Slices

Abstract:
Quantitative label-free imaging with phase and polarization (QLIPP) techniques have garnered considerable attention in various scientific fields, particularly in the realm of biomedical research and medical imaging. By combining phase and polarization information, computational optical imaging approaches enable non-invasive analysis and characterization of samples based on their intrinsic properties without the need for exogenous labels or contrast agents. However, label-free Cell Painting which is a process for predicting the location of protein subcellular organelles often involves dealing with complex tissue structures, such as overlapping or interconnected regions. QLIPP may struggle to delineate such intricate structures. Quantitative label-free imaging techniques rely on the intrinsic properties of the sample to produce contrast, which may be limited in certain cases. With low inherent contrast, it becomes more difficult to segment structures or regions of interest accurately. Therefore, label-free Cell Painting remains particularly challenging for complex and lowcontrast source images. To take full advantage of the great potential shown by QLIPP for label-free Cell Painting tasks, we propose a Swin Transformer based network called ST-Net, and experiments on a mouse kidney tissue slice dataset show that the ST-Net delivers significant performance gains compared to the advanced methods in this field.