A stability-enhanced CycleGAN for effective domain transformation of unpaired ultrasound images

Abstract

As a cost-effective, non-invasive and radiation-free medical imaging modality, ultrasonic imaging is widely used in clinical diagnosis. However, database bias is commonplace among different medical centers. Deep learning based ultrasound image analysis algorithms are usually data driven, rendering high requirements on the uniformity of image datasets. In this paper, we propose a stability-enhanced cycle-consistent generative adversarial network (CycleGAN) method with well detail preservation for the domain transformation to normalize ultrasound images from various medical centers. To stabilize the training process of CycleGAN model, we adopt spectral normalization for 1-Lipschitz continuity to reduce model oscillation. Besides, two time-scale update rule and label smoothing strategy are also utilized to maintain the balance between generators and discriminators for further stability enhancement. Moreover, our method applies skip connections to preserve ultrasound image details and prevent resolution loss during the domain transformation process. Experiments were conducted on clinical thyroid and carotid image datasets acquired from several medical centers. Massive results demonstrate that our proposed model is easier to reach a steady state when training, outstanding 50% from the basic CycleGAN model. Compared with representative algorithms, our proposed method reaches the state-of-the-art performance, with a 11.3% decrease in the mean absolute error and a 9.8% increase in the structural similarity. Hence, our proposed algorithm has a strong capacity of the domain transformation in ultrasound images to reduce the database bias for uniformly distributed datasets. We believe that our method can contribute to the development of the ultrasound image analysis and computer aided clinical diagnosis..