search

yifangao112 / DeSAM

The official repository for DeSAM: Decoupled Segment Anything Model for Generalizable Medical Image Segmentation.
MIT License
113 stars 5 forks source link

Instance segmentation #1

Closed thangngoc89 closed 1 year ago

thangngoc89 commented 1 year ago

Thank you for releasing your works. May I ask if this model is suitable for instance segmentation? Like there are multiple masks of the same object type.

TIA

yifangao112 commented 1 year ago

Hi @thangngoc89,

Yes, DeSAM is suitable for instance segmentation tasks. It can do the same thing as U-Net, though it does not currently support multi-class segmentation. Multi-class segmentation requires a slight modification of the IoU prediction head.

thangngoc89 commented 1 year ago

@yifangao112 thank you for answering. I actually only need a single class instance segmentation. How would I prepare data for such case?

TIA

yifangao112 commented 1 year ago

Hi @thangngoc89,

The format of our raw data follows nnU-Net. you can find the code to convert the dataset (for both biomedical and non-biomedical images) at this link. In addition, it is necessary to modify the code which divides the data into training set, validation set and test set.

thangngoc89 commented 1 year ago

@yifangao112 thank you very much