Significant advances have been made towards building accurate automaticsegmentation systems for a variety of biomedical applications using machinelearning. However, the performance of these systems often degrades when theyare applied on new data that differ from the training data, for example, due tovariations in imaging protocols. Manually annotating new data for each testdomain is not a feasible solution. In this work we investigate unsuperviseddomain adaptation using adversarial neural networks to train a segmentationmethod which is more invariant to differences in the input data, and which doesnot require any annotations on the test domain. Specifically, we learndomain-invariant features by learning to counter an adversarial network, whichattempts to classify the domain of the input data by observing the activationsof the segmentation network. Furthermore, we propose a multi-connected domaindiscriminator for improved adversarial training. Our system is evaluated usingtwo MR databases of subjects with traumatic brain injuries, acquired usingdifferent scanners and imaging protocols. Using our unsupervised approach, weobtain segmentation accuracies which are close to the upper bound of superviseddomain adaptation.
translated by 谷歌翻译