[深度学习从入门到女装]Few-shot 3D Multi-modal Medical Image Segmentation using Generative Adversarial Learning

论文地址:Few-shot 3D Multi-modal Medical Image Segmentation using Generative Adversarial Learning

 

这是一篇使用GANs进行semi-supervised 3D多模态医学图像分割的论文

 

[深度学习从入门到女装]Few-shot 3D Multi-modal Medical Image Segmentation using Generative Adversarial Learning_第1张图片

 

对于semi-supervised任务,本文对GANs的loss进行改进:

[深度学习从入门到女装]Few-shot 3D Multi-modal Medical Image Segmentation using Generative Adversarial Learning_第2张图片

[深度学习从入门到女装]Few-shot 3D Multi-modal Medical Image Segmentation using Generative Adversarial Learning_第3张图片

由于对于K+1个分类情况下,最后一个对于fake的预测是基于softmax的,所以使用K+1个分类loss就显得多余,于是本文只采用K个输出,相当于多softmax进行一个normalized

因此,修改过的loss为:

[深度学习从入门到女装]Few-shot 3D Multi-modal Medical Image Segmentation using Generative Adversarial Learning_第4张图片

 

对于semi-supervised任务来说,如果使用normal GAN中的generator loss 可能会导致不稳定性,因此本文加入了Feature Matching loss在discriminator中,f得到的是discriminator(U-net)中encoding path的倒数第二层的activation结果

 

对于上文讲的FMloss,会带来FM可能只match了first-order的统计,generator最终会进行trivial,

[深度学习从入门到女装]Few-shot 3D Multi-modal Medical Image Segmentation using Generative Adversarial Learning_第5张图片

你可能感兴趣的:(深度学习)