Inquiry Regarding Registration Model Inference for Pre-to-Post Surgical Prediction

[wangnanv5]

Hello, I have a question regarding image registration that I would like to consult with you.
Currently, I possess paired binary images consisting of pre-operative and post-operative CT bone segmentation results for pectus excavatum surgery. My goal is to design a registration-based training framework where the model can generate the post-operative segmentation map solely by taking the pre-operative image as input during inference.
The objective of this experiment is relatively straightforward: to simulate minor deformations specifically in the skeletal structures surrounding the sternum. However, I have observed that all registration models available in MONAI require both a "moving image" and a "fixed image" (reference image) as inputs during the inference stage.
Could you please advise on how I can configure the workflow so that, during inference, the model only requires the pre-operative segmentation as input to predict the post-operative result?
I look forward to your valuable guidance.
Best regards,

[ericspod]

I think what you want is just to apply freeform deformation to an image rather than doing registration. The idea with registration is that you have two images which should correspond in some way, then you predict or estimate the deformation field to get the moving image to match the fixed image. If you have only one image as input then you don't have any correspondence between images to predict, so I think what you want to do isn't registration but image generation.

If you want to start from a pre-operative image and generate a post-operative one (segmentations being an image type), you are looking at an image-to-image problem which you could tackle with generative networks. You would need a dataset of pre-operative images paired with corresponding post-operative segmentations, then with that you can train a network to predict the segmentation. This likely isn't easy and will require a complex supervised or semi-supervised scheme, so you should consult much of the existing literature on predicting post-operative biomarkers from pre-operative imaging. This isn't my area so it's hard for me to advise beyond that, but I don't think registration is the way to go from your description.

[wangnanv5]

Previously, I attempted to accomplish this task using a generative model, but the result was extremely poor. Therefore, I decided to start from the registration aspect to see if it would work. Through your explanation, I learned that this approach is not feasible. Thank you for your guidance.😊