search

m1nhengChen / SOPI

Code for ICASSP 2024 paper"Embedded Feature Similarity Optimization with Specific Parameter Initialization for 2D/3D Medical Image Registration"
17 stars 0 forks source link

RTPI_train about #4

Open ryukyungjoon opened 4 weeks ago

ryukyungjoon commented 4 weeks ago

Hello, Chen

I have a question about learning RTPI. Is there anything I should consider when choosing a training dataset? For example, the learning volume must have the same pixel spacing and be the same size.

weight_l2_norm_rot = weight_l2_norm_trans = weight_l2_norm_trans1 = weight_l2_norm_trans23 = weight_l2_norm = weight_ncc =

Also, ProST_drr_generator and how can I set these parameters?

Thank you.

m1nhengChen commented 3 weeks ago

Hi @ryukyungjoon ,

There are no explicit restrictions on the size of the CT volume or the pixel space of the projected image, but you must ensure that the parameters in ProST are consistent with your X-ray and CT data. Usually those coefficients are obtained by grid search, you may need to try to find the optimal coefficients on your own dataset, as in our data the weight ratio of rotation and translation is 1:1.6.

ryukyungjoon commented 1 week ago

Thank you @m1nhengChen

How to fit the parameter?

RuntimeError: shape '[-1, 128, 64, 64]' is invalid for input of size 32768

my custom dataset ct volume size is (83, 512, 512).