Rough reading DS transunet: dual swing transformer u-net for medical image segmentation

This article uses the twin network , Through different patch_size, Pay attention to the information of different sizes , And then use it TIF Module integration .

So the core innovation points are two points ： First of all , Through different patch_size Focus on information at different scales , second , adopt TIF Fusion information .

TIF The module is shown in the figure , It is to pool the other party's characteristic graph , And then as the tail of the characteristic graph added to the other side after flattening , Conduct transformer Calculation

meanwhile ,swin transformer Part is shown in the above figure ,attention Modules no longer focus on the overall situation , But in two steps , Every step every patch Only a few of the surrounding patch Calculate ,SW-MSA The difference is that it will move the whole image to the left , The disappearing part emerges from the right and below , And then calculate MSA. See this article for details ：SW-MSA