search

icon-lab / SynDiff

Official PyTorch implementation of SynDiff described in the paper (https://arxiv.org/abs/2207.08208).
Other
229 stars 39 forks source link

how to preprocess the original IXI dataset like your format? #13

Closed sunshinelike closed 1 year ago

sunshinelike commented 1 year ago

how to preprocess the original IXI dataset on line like your format?

onat-dalmaz commented 1 year ago

Thank you for your comment! First of all, sequentially select the first 40 subjects from the dataset. Since multi-contrast MRI volumes were unregistered as publicly shared in IXI, T2- and PD-weighted volumes were registered onto T1-weighted volumes in IXI Registrations were implemented in FSL via affine transformation and mutual information loss. In each subject, each imaging volume was separately normalized to a mean intensity of 1. The maximum voxel intensity across subjects was then normalized to 1 to ensure an intensity range of [0,1]. We hope this information helps clarify the preprocessing steps taken for the dataset used in our study.

sunshinelike commented 1 year ago

Thank you very much for your detailed answer. Have your experiment in IXI domain T1 and T2 translation used all data of 662 volumes? In your readme file, if we want translate T1 to T2, they must be registered to the same in both train dataset, test dataset, right? I use the same seed and split method to split the dataset, but get the unmatched T1 and T2 in corresponding dataset. Thank you very much if your could give some advice.

onat-dalmaz commented 1 year ago

No we only used 40 subjects, with a (25,5,10) split for train, validation, and test sets. For validation and test sets, to be able to accurately calculate performance metrics, we have registered T2 volumes to T1 volumes. Since the method is unsupervised, there is no registration requirement for the training set.

sunshinelike commented 1 year ago

Thank you very much for your helpful and quick response. OK, I see. Best wishes.

Luke-Huang-Github commented 10 months ago

@onat-dalmaz Hi, thank you for your contribution. I have some questions about the details of the dataset. (1) In IXI dataset, First of all, sequentially select the first 40 subjects from the dataset,with a (25,5,10) split for train, validation, and test sets. In paper "each subject 100 axial cross-sections with brain tissue were selected". Do you mean the data_train_contrast1.mat shape is (2500, width, height) -> (#images, width, height)? #images = 25*100?