icon-lab / ResViT

Official Implementation of ResViT: Residual Vision Transformers for Multi-modal Medical Image Synthesis
Other
133 stars 27 forks source link

Data pre-processing #20

Open LightA2 opened 1 year ago

LightA2 commented 1 year ago

Hello, the author has the honor to communicate with you, inspired by your work, recently in the study of modal synthesis work, the process encountered some problems, I hope to get your answer! The problem is: when one mode generates another, we need to splice the two modes together. For example, when a T1 mode generates a T2 mode, we splice the slices of the two modes together horizontally, and what happens when we need multiple modes to generate a single mode? For example, the T1 and T2 modes generate Flair modes. Do we need to splice the data of these three modes horizontally together, or is it some other processing method. I sincerely hope to get your answer!

onat-dalmaz commented 1 year ago

In case of task-specific models, to generate FLAIR modality from T1 and T2, first T1 and T2 images are concatenated in channel dimension. Then, networks is fed with this multi-channel input to generate a FLAIR image with only one channel.

For unified models, each modality has a specific channel in a multi-channel input/output framework. For instance: [T1,T2,0]->network->[T1,T2,FLAIR]

I hope this addresses your inquiry.