基于半监督级联网络的高质量医学图像集增强

Enhancement of high-quality medical image set basedon semi-supervised cascaded network

应用于医学图像处理的深度神经网络在训练阶段需要大量高质量图像,因此,在医学图像数据量有限的情况下,网络性能会受到局限。虽然现有的基于生成对抗网络GAN的增广方法能显著增加图像的数量,但是存在合成图临床表征不准确和缺乏多样性等问题。为解决上述问题,提出了一种新颖的基于半监督学习的多输入多分辨率多模板的生成对抗网络。多通道的输入分别为模型的训练提供了基于监督学习和无监督学习的优化目标;多分辨率级联策略降低了直接生成高分辨率医学图像的难度;多类别的参考模板为各通道及各分辨率尺度上的训练提供更准确的真实医学图像临床表征;引入了过渡机制和稠密残差块,提升了模型训练的稳定性。实验结果表明:相较于其他基于GAN的生成模型,该网络模型能生成更高质量、更具多样性的医学图像。

The deep neural network used for medical image processing requiresa large number of high-quality images in the training stage.Therefore,in the case of limited medical image data,the network performance will be limited.Although the existing GAN-based augmentation methods can significantly increase the number of images,there are problems such as inaccurate clinical representation of composite images and lack of diversity.In order to solve the above problems,a novel multi-input multi-resolution multi-template generative adversarial network based on semi-supervised learning is proposed.Among them,multi-channel input provides optimization targets based on supervised learning and unsupervised learning for model training;multi-resolution cascade strategy reduces the difficulty of directly generating high-resolution medical images;multi-category reference template provides more accurate clinical representation of real medical images for training on each channel and resolution scale;the transition mechanism and dense residual blocks are introduced to improve the stability of model training.Experimental results show that compared with other GAN-based generative models,the network model can generate higher quality and more diverse medical images.