Abstract
Existing deep neural networks for histopathology image synthesis cannotgenerate image styles that align with different organs, and cannot produceaccurate boundaries of clustered nuclei. To address these issues, we propose astyle-guided instance-adaptive normalization (SIAN) approach to synthesizerealistic color distributions and textures for histopathology images fromdifferent organs. SIAN contains four phases, semantization, stylization,instantiation, and modulation. The first two phases synthesize image semanticsand styles by using semantic maps and learned image style vectors. Theinstantiation module integrates geometrical and topological information andgenerates accurate nuclei boundaries. We validate the proposed approach on amultiple-organ dataset, Extensive experimental results demonstrate that theproposed method generates more realistic histopathology images than fourstate-of-the-art approaches for five organs. By incorporating synthetic imagesfrom the proposed approach to model training, an instance segmentation networkcan achieve state-of-the-art performance.