dykuang / medical-image-registration Goto Github PK

Hello, may I ask whether the pixel value of some blocks of the image is 0 after the displacement occurs in the process of debugging the code?Thank you very much!

Hello,
i need to ask one question. can you please tell me which function you use for visualization of grid.
That you show in folding effect grid
https://github.com/dykuang/Medical-image-registration/blob/master/CheckJac_short.png
shown in this pic
Kind regards
Hafiz

I find that you calculate the dice based on the label list, and then average each part of a sub-region.Why not first fuse all labels into the five rois, and then calculate the dice?

It's like a math question,I thought a lot but can't find the answer.

def Get_Ja(displacement):

'''
Calculate the Jacobian value at each point of the displacement map having

size of b*h*w*d*3 and in the cubic volumn of [-1, 1]^3

'''

D_y = (displacement[:,1:,:-1,:-1,:] - displacement[:,:-1,:-1,:-1,:])

D_x = (displacement[:,:-1,1:,:-1,:] - displacement[:,:-1,:-1,:-1,:])

D_z = (displacement[:,:-1,:-1,1:,:] - displacement[:,:-1,:-1,:-1,:])



D1 = (D_x[...,0]+1)*( (D_y[...,1]+1)*(D_z[...,2]+1) - D_z[...,1]*D_y[...,2])

D2 = (D_x[...,1])*(D_y[...,0]*(D_z[...,2]+1) - D_y[...,2]*D_x[...,0])

D3 = (D_x[...,2])*(D_y[...,0]*D_z[...,1] - (D_y[...,1]+1)*D_z[...,0])

return D1-D2+D3

Thanks for the code.

I am trying to save the checkpoints after every epoch by the following.

from keras.callbacks import ModelCheckpoint
mc = ModelCheckpoint(outpath+'SDN3d_weights_TVS16_{epoch:02d}.ckpt', save_weights_only=True, verbose=1, period=1)
history = sdn.fit_generator(gen_train, steps_per_epoch=2, callbacks=[mc], epochs=2, verbose=1, validation_data=gen_test, validation_steps=len(f_img_v)/batch_size)

the above code generates the checkpoints file but when I try to load them in the eval.py by sdn.load_weights(r'path_to_file'), there is error

"ValueError: You are trying to load a weight file containing 2 layers into a model with 18 layers."

Thanks in advance

Dear @dykuang,
first of all thanks a lot for sharing your code.
I read the two papers in arxiv and I am interested in the problem of the negative Jacobian in the deformation field. Looking at your code I did not really understand the way you calculate the Jacobian and the related function Get_Ja(). Can you please explain me how you

Calculate the Jacobian value at each point of the displacement map

Dear dykuang,
This repository helps me a lot, but there is something confused me a lot. The question is how to preprocess the data? Can you explain this to me?
Thanks

Hi! Dear Author. I'm studying your paper<FAIM – A ConvNet Method for Unsupervised
3D Medical Image Registration>.
I'm a little confused about your loss function.

1. What does -1 in the little box mean?
   

Looking forward to your reply, thank you

Hi,Dear Author,Your code has helped me a lot.
I noticed that the 'Get_Jac()' in 'Utils.py' appaers '+1'. So...what does this '+1' do.
Just like this picture

Is there any instruction for fusion the labels into the five ROIs?

Dear dykuang,
Your work has inspired me a lot, but there is some code I got confused.
1. In train.py, val_files = ['{:03d}'.format(i) for i in range(0, 20)]
Which data subset is it(val_files) in mindboggle? I guess it should not be OASIS-TRT-20, because this subset is used to evaluate the network as test set.
2. In addition. In voxelmorph, when Unet is down-sampling, the image size will be reduced by a factor of 1/2. If the data set is cropped to 144x180x144, when reproducing voxelmorph, the value 180 will produce fractions during down-sampling,which is automatically rounded down by the computer. Then the size of the image after down sampling will not match the up-sampling, wouldn't it make it impossible to train?
Thanks

Hello, I am very lucky to see your research.If I only want to register a pair of 3D images, which script should I run?

Hello, dykuang:

I'm implementing your vis_grid function in visual.py, The registration field I used is (31,31,3),
and i use the command line vis_grid(flow[:, :, ::2], res=1, direct=2),after that, i got a likely wrong result. The result is:
444