Echocardiography Segmentation Based on Cross-modal Data Augmentation Method

Deep learning algorithms have been widely applied in the field of medical image analysis. However, it is very difficult and time-consuming to obtain a large number of accurate samples and their corresponding labels. Deep neural networks can be difficult to train without a large number of samples, which may lead to unsatisfactory performance. Currently, synthesizing medical image samples is an effective solution to this problem. One of the commonly adopted methods for synthesizing labeled ultrasound images is simulation-based approach, such as Field II software. Although these simulation-based methods have demonstrated acceptable performance, they are not only time-consuming, but also require expert experience to fine-tune parameters. Therefore, we propose a CycleGAN-based method to synthesize annotated ultrasound images from computed to-mography (CT) images. Our proposed method is validated on the left ventricle segmentation task. Experimental analyses demonstrate that our proposed annotated sample generation method outperforms the general CycleGAN method in the context of segmentation results.