Revisiting Self-Generating Simple Figure Patterns for Learning Microscopy Image Segmentation

Learning microscopy image segmentation models often requires costly large-scale annotated datasets. Collecting microscopy images demands expensive specialized equipment, and annotating them requires expertise. While research has sought methods to reduce the dependency on annotated data, challenges associated with the collection and annotation of specialized datasets remain prevalent. In this study, we show that using simple figure patterns can effectively facilitate the training of a two-class segmentation model designed to distinguish foreground from background in microscopy images. Our approach leverages a generator-mixer module to synthesize training data from predefined figure patterns, which is then used to train a deep learning model for image segmentation. Evaluation demonstrates that this approach effectively segments cellular and particulate regions within various microscopy images, showcasing both versatility and tunability, and highlighting the potential of synthetic data as a viable resource for training segmentation models in microscopy applications.