search

bowang-lab / MedSAM

Segment Anything in Medical Images
https://www.nature.com/articles/s41467-024-44824-z
Apache License 2.0
3.03k stars 419 forks source link

Why MedSAM2 did not use transfer learning for the same 1.5m image-mask pairs as in MedSAM1? #311

Closed dzzhang96 closed 1 month ago

dzzhang96 commented 2 months ago

In the paper, MedSAM2 only fine-tuned the model by transfer learning on abdominal organ segmentation datasets based on SAM2-T. Why not use the same 1.5m datasets as in MedSAM1? Thank you!

JunMa11 commented 2 months ago

Hi @dzzhang96 ,

The main reason is that we currently don't have enough computing.

Another important reason is that we didn't find SAM2 has too many advantages for image segmentation. We're working the video fine-tuning code but this is very challenging compared to image fine-tuning. Stay tuned!