Domain adaptation in nuclei semantic segmentation

Medical image segmentation has long been suffering from the lack of datasets since labelling pathological data is laborious work and requires specialized skills, which could only be done by professional doctors, especially when it comes to nuclei semantic segmentation. Besides, due to the fact that the domain gap inevitably exists between different datasets, which could be caused by diversified staining methods or the heterogeneous appearance of different tissues, it is almost impossible to get labelled data under all circumstances. This paper applies domain adaptation as an effective and efficient method to align two domains in latent feature space. We experiment on both IoU and Excepted Calibration Error (ECE), an indicator mostly used in biomedical segmentation to evaluate our work. In two domain adaptation tasks, i.e., TNBC and MoNuSeg, we proved that by exchanging the low frequency of two styles of the datasets, can Fourier Domain Adaptation (FDA) successfully achieve a considerable increasement of 1% and 2.29% higher than simply using source images to train with U-net in the target dataset.