search

wuzijian1997 / fine-tuning-MedSAM

Apache License 2.0
3 stars 0 forks source link

Labels for Finetuning #1

Open Testbild opened 4 days ago

Testbild commented 4 days ago

Dear Wuzijian1997,

I would like to use the medSAM model for my dataset and finetune for my pathology. Do I need annotated labels for that, and if so, how many do I approx. need?

Initially my thought was, that I do not need any labels for the medSAM. When I use the GUI version it also segments pretty good in the boxes I provide, but currently I have no way of generating the boxes automatically and I would like to run it on a large dataset to segment anything by itself.

I am just a bit confused about the fine tuning explanation in the repo, maybe you can clarify that for me.

Thank you and best regards

wuzijian1997 commented 4 days ago

Hi, you do need labels to fine-tune SAM/MedSAM/xxxSAM...

There might be some solutions for you. SAM 2 can deal with videos. You only need to provide prompts in the initial frames to segment the first frame well. You can try to use SAM 2 to generate labels for fine-tuning. But you still need labels to fine-tune SAM.

Feel free to reply if you have further questions!