Abstract
Convolutional neural network (CNN), in particular the Unet, is a powerfulmethod for medical image segmentation. To date Unet has demonstratedstate-of-art performance in many complex medical image segmentation tasks,especially under the condition when the training and testing data share thesame distribution (i.e. come from the same source domain). However, in clinicalpractice, medical images are acquired from different vendors and centers. Theperformance of a U-Net trained from a particular source domain, whentransferred to a different target domain (e.g. different vendor, acquisitionparameter), can drop unexpectedly. Collecting a large amount of annotation fromeach new domain to retrain the U-Net is expensive, tedious, and practicallyimpossible. In this work, we proposed a generic framework to address thisproblem, consisting of (1) an unpaired generative adversarial network (GAN) forvendor-adaptation, and (2) a Unet for object segmentation. In the proposedUnet-GAN architecture, GAN learns from Unet at the feature level that issegmentation-specific. We used cardiac cine MRI as the example, with threemajor vendors (Philips, Siemens, and GE) as three domains, while themethodology can be extended to medical images segmentation in general. Theproposed method showed significant improvement of the segmentation resultsacross vendors. The proposed Unet-GAN provides an annotation-free solution tothe cross-vendor medical image segmentation problem, potentially extending atrained deep learning model to multi-center and multi-vendor use in realclinical scenario.