Determining optimal hyperparameter settings for abdominal CT image registration

[MaKh001]

Dear ANTs developers and experts，

I'm currently working on enhanced abdominal CT image registration and have found that the default parameters do not yield optimal registration results. Although the registration can be improved via key parameters tuning, the optimal configuration for specific cases varies and I found no clue to infer the optimal configuration from image pairs intuitively other than performing a grid search.

Here's the keys parameters that affect the registration:

-Registration Method: antsRegistrationSyN
-Key Parameters:
-r: radius for cross correlation metric used during SyN stage (default = 4), varied from 0 to 8 in my cases
-s: spline distance for deformable B-spline SyN transform (default = 26) , varied from 20 to 80 in my cases

If any professionals have experience in abdominal CT image registration and parameter tuning, could you offer insights or suggestions on refining the parameters? Any advice on key parameters or general strategies for improvement would be greatly appreciated.

Thank you for considering my question.

[ntustison]

This isn't descriptive enough to diagnose the problem. For example, we would also need to understand how you're assessing "optimal registration results" and to actually see the data and results ourselves, or at least a few examples.

[MaKh001]

Thank you very much for your response. We segmented images of multiple organs, including spleen, kidney, gallbladder, liver, pancreas, adrenal glands, blood vessels, and bones. After using specific hyperparameters for registration, the average dice coefficient of multiple organs is calculated, and the corresponding hyperparameter when reaching the maximum value is regarded as the optimal hyperparameter. I put the relevant NIfTI files of the registration results of the default hyperparameters and optimal hyperparameters on Google Drive. You can download them using the following link:
https://drive.google.com/drive/folders/1S9ilYpq7k5tnoYWupMblrDj1-rGk7K-I?usp=sharing

In addition, I put the calculation results of these two examples in the figure below.
case1


case2


Once again, sincere thanks for your attention and assistance.

[ntustison]

I don’t have access to your data but I can provide some advice even before looking at the images.

Given the large number of parameters to be optimized during image registration, it’s no surprise that individual subjects (and organs) will exhibit a wide range of values for a given evaluation metric. That’s why image registration competitions and other evaluations tend to look at a variety of evaluation metrics over several subjects. And even then, those evaluation metrics might not be sufficient to capture what is actually needed for a specific study. My advice would be to, starting with the default settings in “antsRegistrationSyN.sh” or “antsRegistrationSyNQuick.sh”, actually look at the images and determine where alignment needs systematic improvement over several images for your particular study starting with the default settings. You can then ask follow-up questions here for further guidance. This is much a more principled approach than doing a brute-force grid search where some of the parameters that you’ve included don’t make sense, from an experienced developer’s perspective.

[MaKh001]

Thank you for your advice. I apologize for any inconvenience caused by my unfamiliarity with the sharing process. I can confirm that access has now been successfully granted. Here is the link.
https://drive.google.com/drive/folders/1S9ilYpq7k5tnoYWupMblrDj1-rGk7K-I?usp=sharing

I appreciate your insightful advice and the inspiration you've provided. I'll take some time to carefully consider your suggestions and will provide a detailed response later. Thank you for your understanding, and I'll get back to you as soon as possible.

[ntustison]

Thanks for making them available. However, it looks like the images might not have fully uploaded as I'm only seeing the lower half (see below for an example) even though the segmentations seem to represent the full view. If you have comprehensive segmentations for each image, you might want to augment your registration set-up by using each segmentation region as a separate image metric. So, for example, with antsRegistrationSyN.sh, suppose you had a left lung segmentation and right lung segmentation as separate binary images along with your ct image for both the fixed and moving image sets, you could do something like

$ antsRegistrationSyN.sh -d 3 -f fixed_ct.nii.gz -m moving_ct.nii.gz \
                                                     -f fixed_left_lung.nii.gz -m moving_left_lung.nii.gz \
                                                     -f fixed_right_lung.nii.g.z -m moving_right_lung.nii.gz \
                                                     ...

You should be able to set that up for any number of your segmentated organs.

[MaKh001]

Thank you for your advice and the provided code. However, I'd like to explain that our images are complete. We cropped the abdominal CT images to focus on the upper abdomen for faster computation while retaining essential information.

I apologize for any confusion caused by the unclear presentation of our research purpose and results earlier, potentially leading to some unnecessary time spent on your part. I have revised our results to provide a clearer presentation, as shown in the updated figure.

Due to the observed variations in the optimal parameters for each pair of abdominal CT images, we aim to predict the best registration parameters from resolution or other image features to enhance the precision of the image registration. Our study specifically focuses on parameters r and s, which we consider to be meaningful contributors, and our research has revealed their impact on the registration outcomes. Additionally, we aim to enhance the persuasiveness of our results through quantitative assessments, with the average Dice coefficient calculation across multiple organs being a commonly used method, so we chose it.

Your previous advice has been instrumental in our search for optimal parameters. However, we also hope you can offer insights into determining these parameters through image features.

Thank you sincerely for your prior guidance. We greatly appreciate your support and would be grateful for any additional insights you may provide.