基于生成对抗网络的乳腺MRI图像生成

Breast MRI image generation method based on generativeadversarial network

乳腺癌磁共振成像(nuclear magnetic resonance imaging,MRI)数据由于不同医院采集方式不同、设备不同或病人等自身原因,会存在同一病人不同序列缺失的问题。目前主流的图像生成对抗网络Pix2Pix和Cycle-consistency是医学图像生成的两种主要模式,这类方法要求不同MRI序列数据配对出现,难以处理存在缺失的数据,此外,该类方法往往关注整幅图像的生成质量,缺少对疾病诊断更有价值的病灶区域的生成质量的监控。针对以上问题,该文受配准网络(RegGAN)自适应对准图像空间分布的启发,设计了一种新的基于特征增强的双注意力配准生成对抗网络DA-RegGAN。该网络在生成器中引入卷积注意力模块,使网络更注重病灶的学习;在判别器中添加梯度正则化约束,主要解决网络训练不稳定容易出现模式崩溃的现象,使网络生成包含更清晰的病灶细节全局图。该文在1697幅乳腺数据上开展消融实验、不同图像生成算法间的对比实验、肿瘤分类实验,进一步验证了方法的有效性。与原始RegGAN比,全局图像生成质量和局部病灶图像生成质量均得到提升,局部图像质量较原始PSNR提升了0.518,SSIM提升了0.021;全局图像质量较原始PSNR提升了0.584,SSIM提升了0.020。

Breast cancer MRI data often have the problem of different sequences missing from the same patient due to different acquisition methods,different equipment or patients and other self-inflicted reasons in different hospitals.These methods often require different MRI sequences to appear in pairs,making it difficult to handle missing data.In addition,these methods tend to focus on the quality of the entire image,and lack the ability to monitor the quality of the focal area,which is more valuable for disease diagnosis.In this paper,we design a new feature-enhanced dual-attention aligned generative adversarial network,DA-RegGAN,inspired by the adaptive alignment of image spatial distribution in RegGAN.The gradient normalization constraint is added to the discriminator,which mainly solves the phenomenon that the network is unstable and prone to pattern running,and enables the network to generate a global map containing clearer lesion details.In this paper,ablation experiments,comparison experiments between different image generation algorithms and tumour classification experiments are carried out on 1697 breast case data to further validate the effectiveness of the method in this paper.Compared with the original RegGAN,the global image generation quality and local lesion image generation quality were both improved,with the local image quality improved by 0.518 and SSIM improved by 0.021 compared with the original PSNR;the global image quality improved by 0.584 and SSIM improved by 0.020 compared with the original PSNR.