search

hsokooti / RegNet

Nonrigid image registration using multi-scale 3D convolutional neural networks
Apache License 2.0
192 stars 50 forks source link

patch concatenate #9

Closed cs123951 closed 5 years ago

cs123951 commented 5 years ago

Hi, thank you for that your code have helped me a lot. However, I am confused about some of your ideas, especially about patch.

  1. From the function “next_sweep_patch()” in functions.reading.real_pair, it seems that you get patches from image pair, and then concatenate these patches to form DVF. But this is easy to generate wrong correspondings between adjacent DVF. i.e. It's obvious that there are traces between patches in the final DVF. Like this: image

How to solve this problem?

  1. The input size in Function.RegNetModel.decimation3 is (155,155,155) and the output size is (27,27,27,3). Then the output is the DVF for the center part (27,27,27) in (155,155,155)?

Thanks a lot if you could answer my doubts.

hsokooti commented 5 years ago

Hi Joy Lu,

Thank you for your interest in our project.

  1. About the stitching issue, is the attached DVF obtained from the decimation3 network? Unfortunately, I do not have any DVF samples from the decimation3 network in the meantime. However, I checked the results of decimation4 network and apparently there is no issue in the stitching process: Sample DVF_s The input size of this network is 105 and the output size is 21. It is possible to increase the input/output size with a multiple of 4 at the test time.

  2. Your point is completely true.

cs123951 commented 5 years ago

Thank you for your rapid reply! The stitching issue is not from decimation3 network but from my self-designed network. But your answers is a great help to me. Thank you again.