Image registration is a typical problem in image processing and technical difficulties in the study, it is an object for comparison or fused image of the same object acquired under different conditions, for example, from a different image acquisition devices will, at different times from different the shooting angle and so on, and sometimes need to use for image registration problems of different objects. More specifically, for a set of two image data sets, by finding a spatial transformation to one image to another image map, so that in both figures the same spatial position corresponding to the match point up to achieve information integration.
This article was published in 2018 , the authors used CNN to achieve a non-supervised medical image registration accuracy and STOA flat, but due to the unsupervised speed increased dramatically. Contributed three articles:
- Proposed a solution based learning, such as do not need to acquire during the training ground truth corresponds or anatomic landmarks and other information;
- Propose a parameter shared across the population CNN function, function by assessing registration;
- Parameter optimization cost function may use various methods to accommodate a variety of tasks;
Network Architecture Figure:
输入为160 × 192 × 224×1的两张单通道灰度图（moving和fixed）叠加成160 × 192 × 224×2。通过CNN学习函数g，CNN的输入即φ配准域，p是体素，φ(p)是图像中的具体位置，F(p)和M(φ(p))确定了相似的解剖位置。
文章定义了空间转换函数，使用差值的形式将配准域转换为最终的配准图：
网络通过梯度下降进行优化，定义的损失函数是常规配准的能量函数上进行改进，常规的能量函数做代价函数如(1)(2)，其中Lsim(·, ·)定义了原图和配准图的相似度，Lsmooth是为了配准图变换得更加平滑而增加的惩罚项。
本文的损失函数定义为：
其中CC为互相关函数：
测试评价指标为常见的Dice Score:
CNN article used is similar Unet structure:
