Lien DOI – 10.1109/ISBI48211.2021.9433864
Domain adaption is a tool to fit models trained on a source dataset to characteristically different target samples. During training, such methods typically seek to minimize the segmentation loss, in conjunction with a penalty function for domain misalignment. State-of-the-art solutions are sensitive to heuristically chosen hyper-parameters that dictate the proportion of the two cost functions. We address this issue by introducing a novel strategy for hyper-parameter estimation via min-max optimization of the deep neural network’s loss function. Our solution provides an analytical expression for the model hyper-parameters, which are iteratively estimated during training. Experimental evaluation on both synthetic data and microscopy images of ameoboid cells Entamoeba histolytica attest to the effectiveness of our solution for deep domain-adapted segmentation in bioimaging.