基于深度学习的医学影像高效生成方法研究

Research on medical image generation method based on deep learning

目的:提出一种基于深度学习的医学影像高效生成方法,以解决医学影像数据获取困难、患病样本分布不均匀的问题。方法:将作为CycleGAN输入的2种真实图像进行图像风格迁移,在训练过程中CycleGAN学习2种图像的图像风格,并将这2种图像的图像风格进行转换,输出不同于原图像风格的图像。以肺部CT图像和眼底图像为例,分别对该方法的样本风格转换能力和患病样本生成能力进行测试。结果:经过训练后,该方法能够将肺部厚层CT图像转换为薄层CT图像,并能生成大量质量较高的眼底患病图像,且生成图像耗时短。结论:该方法可以高效生成医学影像,为临床研究以及相关人工智能模型训练提供了有力的保障,并且可为医学人工智能产品的泛化能力提供测试支撑。

Objective To propose a deep learning-based method for efficient medical image generation to solve the problems of difficult medical image data acquisition and uneven distribution of diseased samples.Methods Two kinds of real images were input to CycleGAN for image style migration,and CycleGAN learned the image styles of the two kinds of images by training,conducted image style conversion and output images different from the original style.The method's sample style conversion ability and sample generation ability were tested respectively with lung CT images and fundus images taken as examples.Results After training the method proposed could convert lung thick-slice CT images into thin-slice ones and generate a large number of high-quality images of fundus disease in short time.Conclusion The method generates medical images effectively,facilitates clinical research and related AI model training,and provides testing support for the generalization ability of medical AI products.