Closed zzx332 closed 4 years ago
This repo has not been tested for cross-modality registration purposes. However, if you want to do that, the first step should be normalizing two modalities (as well as train and val) into similar intensity distributions, or it would be bad results (because of the similarity loss).
Thank you for your reply! The auto_liver_mask preprocess in demo.py don't work on my ct dataset,can i crop area around the target object by manual?
Yes, a rough crop is fine.
In order to register MR to CT.I try to change similarloss to miloss(Mutual information).loss=-miloss+0.1reg_loss Here is my result.(traindata:35ct,35mri,random combination). Mi_loss:https://user-images.githubusercontent.com/45729525/81144615-08e8b400-8fa7-11ea-8c3e-e33d36cf8190.png smooth_loss:https://user-images.githubusercontent.com/45729525/81144674-2584ec00-8fa7-11ea-9b95-d844c5842598.png It looks like miloss is not converging.(decreases too much,from -0.02 to -0.35) smooth_loss first decreases and then increases. Here is dice on the test dataset.It looks ok, but the warped image is bad. val_dice:https://user-images.githubusercontent.com/45729525/81145967-f58b1800-8fa9-11ea-88c4-2643784840e6.png img_fixed:https://user-images.githubusercontent.com/45729525/81146587-2ddf2600-8fab-11ea-9aca-071f6256059e.jpg img_moving:https://user-images.githubusercontent.com/45729525/81146583-2a4b9f00-8fab-11ea-962c-4ec433ff3cd6.jpg img_warped:https://user-images.githubusercontent.com/45729525/81146166-531f6480-8faa-11ea-8130-bada6792eb4f.jpg Is it caused by my loss function setting?
This is actually an active research topic... So it is not surprising that some losses may or may not work for this setting.
Ok thanks for your answers!
The following is the data after processed. train dataset is 35 ct. valid dataset is 5. final dice score is 0.5
Is there any problem? Is the intensity difference between the images too large? I want to registration between ct and mr.What process should i do with mri image?