how to keep the resolution

[darrylbobo]

Hi, I may misread or miss some details of the paper.
But from the network, the input spatial size is 272727, the output is 999. So how do you generate the segmentation results in original size?

Thanks!

[josedolz]

Hi @darrylbobo

we sample the input patches (of size 27x27x27) in a way that the predictions, of size 9x9x9, are spatially near to each other. This means that some regions in the input samples overlap between them. For example, the first input patch could be at (0,0,0) and the next one sampled at (0,0,9).

Jose

[darrylbobo]

@josedolz Thanks a lot for the reply.

But I am still a little confused.
For the input patch covering from 0,0,0 to 26, 26, 26 in the original volume, the corresponding ground truth segmentation will also be from 0,0,0 to 26, 26, 26. But the prediction from the model results in a volume which has the spatial size of 9, 9, 9. Then, my question is how to correlate each of 9 x 9 x 9 pixels in prediction with each of 27 x 27 x 27 pixels in the ground truth?
In other words, in the equation 4 of paper 1712.05319.pdf, the y^s_v is not one to one mapped to p^v (x_s).

Maybe you can refer me to the corresponding code. Thanks!

[josedolz]

Hi @darrylbobo

The ground truth is cropped according to the predicted patch, not the input. If you have an input patch of size 27x27x27 after all the convolutions you will end up with a patch of size 9x9x9, located at the center of the input patch, because you get rid off 9 voxels per side. So, basically, what you have to do for the GT is to take the center of the patch (i.e., the voxel 13,13,13) and crop the GT 9 voxels around (from 9,9,9 to 17,17,17). This GT is the one used in eq 4 of the paper, which has the same size as pˆv(x_s).

Jose

[josedolz]

For the code, the sampling process is here:

SemiDenseNet/src/LiviaNet/Modules/IO/sampling.py

Line 188 in 9a578b5

# Extract samples from computed coordinates