A Generative Adversarial Network for Synthesizing Medical Images

Abstract

The invention discloses a generative confrontation network for synthesizing medical images. The topology of the generator is: with 8 levels, the level of the level is the logarithm of the base 2 of the resolution of the image learned at that level; the resolution of the image learned by the lower level level is doubled and the transition to the high level level learning is smooth The resolution of the image; the feature map of the convolutional block is the feature map with self-attention. The topology of the discriminator is opposite to that of the generator. The resolution of the image learned by the high-level layer is half-smoothly transitioned to the resolution of the image learned by the low-level layer; the lowest level of the discriminator also includes batch standard deviation; the generator and the discriminator Use the loss function of WGAN‑GP; initialize the weights using a standard normal distribution, and scale the weights at runtime. Compared with GAN networks that generate large resolutions such as LapGAN and StackGAN, most of the iterations in the training process of the present invention are completed at low resolutions, which greatly speeds up the generation speed on the premise of ensuring the quality of the synthesized images.

Description

technical field [0001] The invention relates to the medical image field of computer vision, in particular to a generative confrontation network for synthesizing medical images. Background technique [0002] With the improvement of computing power and the sharp increase in the amount of data in various industries, artificial intelligence has achieved rapid development. In 2014, Hinton's doctoral student lan Goodfellow proposed the GAN network (Generative Adversarial Networks), and then the improvement of GAN was in full swing in academia, and there is still a lot of room for development. [0003] With the continuous research on GAN, various fields try to combine GAN with the work in the field. Since 2017, the application of GAN in the field of medical images has risen sharply, in which image synthesis, denoising, reconstruction, segmentation, detection and classification have become the main directions of GAN in the field of medical images. Supervised deep learning is the c...

AI Technical Summary

Problems solved by technology

On the other hand, due to the protection of patient privacy, public high-quality data sets in the medical field are very scarce, which limits the development of deep learning in the medical field.

Images