The ending of summer immersion
In the final week of summer immersion, I also finished a project named neural pre-processing. Neural preprocessing is of great importance for various downstream problems like brain anatomy segmentation, brain template generation and etc.
Basically, neural pre-processing includes two parts: image translation (intensity normalization) and image alignment (spatial normalization). The intuitive solution for this problem is using two neural networks. One predicts the intensity translation and the other do the job which register the potential images to templates. However, directly sequentially apply two networks are problematic, as it turns out that image translation network isn't prone to preserving original image's geometry.
To solve the issue, I proposed a method to guarantee the geometry preserving property of the translation network. The idea is that the network instead of generating the image direct it generate mapping which is applied to the input image.
Comments
Post a Comment