when I run the Train_LapIRN_diff_resize.py

[yoonguusong]

this is the result that I got after Train_LapIRN_diff_resize.py and save the new model

and Test_LapIRN_diff_resize.py with changing model directory into new saved model which is .pth

I don't know why I got this result.

I also got these kind of result when I run the other Registration code as well.

is it normal "Failed" result?

but I just run the same Train code and run the same Test code.

could you give me any advise for this?

[cwmok]

Hi @yoonguusong,

I think your training is collapsed. Which dataset do you use for training?

I found that the eps value of NCC similarity matters. You may set the default eps value for the NCC to 1e-4 and re-training the model again. eps value is in https://github.com/cwmok/LapIRN/blob/f2b25e22914582b289304dbd928790d079d52887/Code/miccai2020_model_stage.py#L1210C23-L1210C23.

[yoonguusong]

@cwmok i used the dataset which refered in the code, which is neurite_oasis/neurite-oasis.v1.0

[yoonguusong]

This is range_flow = 0.0

This is range_flow=0.1

This is range_flow=0.2

This is range_flow=0.3

I just change the range_flow value into diverse value,
and the result is like this.
As you can see here, as range_flow is getting bigger, warpped image is getting move to the left side.

Is it reasonable result?
Is there any reason why this move into left side?
if this is right, then could you please let me know why this work like this?
and after I see this result, I wonder why every Spatial Transformer Network use this Range_flow value into STN.
could you give me some advise?
Thank you.

[cwmok]

Hi @yoonguusong,

Your result is not normal.
The problem is not related to "range_flow". The "range_flow" forces every level of the model to output a normalized displacement field that is within the range [-range_flow, range_flow]. It should remain the same throughout the training and testing phase.

Since you are using the same dataset as mine and you have faced similar issues in other registration codes, I guess the issue may caused by your machine's environment. The problem could be solved by changing the Pytorch to a different Cuda version.

Second, as I mentioned, I found that the eps value of NCC similarity matters. You may set the default eps value for the NCC to 1e-4 and re-training the model again. eps value is in https://github.com/cwmok/LapIRN/blob/f2b25e22914582b289304dbd928790d079d52887/Code/miccai2020_model_stage.py#L1210C23-L1210C23.

Hope it helps.