《中国医学装备》杂志社

目的：基于生成对抗网络的智能超声诊断仪自动化修复算法，提升超声诊断仪成像质量，为临床诊断提供影像学依据。方法：数据源自烟台毓璜顶医院临床采集的甲状腺超声影像数据集，共包含5 826张标注的超声结节图像。基于生成对抗网络的判别器判断受损的真实与虚假状态，构建成像自动化修复模型，对关键帧进行迭代处理，按照平滑关系对图像进行分割，并选取具有代表性的6张甲状腺超声诊断仪成像图像，以确定像素值的集中分布状态，考虑损失函数的平均绝对误差，计算修复拟合图像的数值分布权重。基于生成对抗网络的的生成器，结合多尺度图像结构进行像素值的集中分布状态综合预测，调整修复图像边界的连贯性，通过处理复杂的对抗损失，完成智能超声诊断仪成像自动化修复。结果：在6张测试集(I1～I6)不同图像下，所提算法应用后的修复结构相似性指标(SSIM)平均值为0.87，未出现组织边界形态异常或结构失真问题。结论：基于生成对抗网络的智能超声诊断仪自动化修复算法应用后，自动化修复效果好，显著提升了智能超声诊断仪成像的质量与可靠性。

Objective: To explore automatic repair algorithm based on generative adversarial network for imaging of intelligent ultrasound diagnostic instrument, so as to enhance image quality of the imaging of ultrasound diagnostic instrument. Methods: The data originated from dataset of ultrasounic images for thyroid, which were clinically collected at Yantai Yuhuangding Hospital. A total of 5,826 annotated images about ultrasound nodule were included. Based on the discriminator of generative adversarial networks, the true and false states of damage was determined, and automatic repair model for imaging was constructed. The key frames were iteratively processed, and images were segmented according to smoothing relationships. The images of ultrasound diagnostic instrument were randomly selected to determine the state of concentration and distribution of pixel values. Considering the averagely absolute error of the loss function, the numerical distribution weights of the fitting images were calculated and repaired. Based on the generator of generative adversarial networks, the multi-scale structures were combined to comprehensively predict the state of concentration and distribution of pixel values, and adjust the coherence of repaired boundaries. The automatic repair for imaging of intelligent ultrasound diagnostic instrument was completed through processed complex adversarial losses. Results: Under different images of I1-I6, the average value of repaired Structural Similarity Index(SSIM) was 0.87 after the proposed algorithm was applied, and there was not problems about abnormal morphology of tissue boundary or structural distortion. Conclusion: After the automatic repair algorithm based on generative adversarial network for intelligent ultrasound diagnostic instrument is applied, the automatic repair effect is better, which can significantly improve the quality and reliability of imaging of intelligent ultrasound diagnostic instrument.