Semantic information-guided attentional GAN-based ultrasound image synthesis method
Shimeng Shi, Hongru Li, Yifu Zhang, Xinzhuo Wang · Biomedical Signal Processing and Control · 2024
Ultrasound (US) imaging is widely used in clinical practice for the diagnosis and treatment of various diseases. Supervised learning algorithms for US image analysis typically require large-scale labeled data. However, the acquisition of such datasets has always been a challenge in the field of medical ultrasound. We propose a generative adversarial network-based synthesis method guided by semantic prior information and attention for synthesizing realistic US images. In this method, the US image-rich semantic structural information is utilized as prior knowledge, the texture feature is introduced to form semantic labels for guiding the generator synthesis, and the detailed background texture structure is enhanced while considering lesion information. A channel attention module is designed to selectively emphasize important channel information features, which integrates the correlation between all channel feature maps and reduces the loss of high-frequency semantic information in the input labels. Furthermore, we propose a regularization feature loss to punish the deep feature difference between real and synthetic images, retain the high-level semantic features in the source image, and further improve the synthesis quality. A large number of experiments were conducted on two challenging datasets to validate our method. Comparative analysis, ablation experiments, customized synthesis, and segmentation tests all demonstrated that our method can synthesize US images with more detailed structural information and fine contours.