Exploring Deep Learning Domain Adaptation performance: from Covariate shift to Wasserstein and beyond

This thesis investigates the efficacy of Batch Normalization Adaptation, an unsupervised Domain Adaptation technique, in the context of image segmentation of mitochondria utilizing a UNet model. The primary objective is to empirically examine the performance of Batch Normalization Adaptation across different source and target datasets, to potentially predict its effectiveness in an unsupervised manner. Here, we illustrate the key findings derived from the pursuit of this objective.