This post is about how to combine patches to obtain the whole ultrasound simulation. The neural network training procedure is described the below post: https://ufuksoylu.web.illinois.edu/2019/07/21/ulm-through-dl-7/ . In the training, input patches were 64 pixels * 64 pixels and output patches were 128 pixels * 128 pixels.
In the inference part, we used 512 pixels * 512 pixels input patches and obtain 1024 pixels *1024 pixels output patches.
Field II simulation’s size is 3247 pixels * 990 pixels on high resolution grid. On input(low resolution) grid, its size is 1624 pixels * 495 pixels. However, we used only 2560 pixels * 990 pixels on high resolution grid because Field II produces some heavy unrealistic artifacts for very near patches.
For combination purpose, three overlapping patches were enough to obtain the whole filed of view. In overlapping regions, we used both reconstructions by linearly adding them. When overlapping area is closer to edges, we used small coefficients since CNNs might have edge artifacts.
Another point is that there are 495 pixels in x direction on low resolution grid. But we needed 512 pixels(so that it is a power of 2). Basically, we zero padded those patches. When we get the output patches, since there are 1024 pixels in x direction, we truncated to 990 pixels.
Here is an example of whole field of view reconstruction:
Q1 per bubble is 1.014441
Q2 per bubble is 0.063079