search

bowang-lab / MedSAM

Segment Anything in Medical Images
https://www.nature.com/articles/s41467-024-44824-z
Apache License 2.0
3.05k stars 421 forks source link

Question about the IOU loss (2) #308

Closed d4nh5u closed 1 month ago

d4nh5u commented 2 months ago 
          no. we didn't optimize it during fine-tuning.

Originally posted by @JunMa11 in https://github.com/bowang-lab/MedSAM/issues/291#issuecomment-2286929055

Hello, I would like to ask why IoU loss was not optimized in this study. What factors were considered?

JunMa11 commented 2 months ago

hi @d4nh5u ,

it can be easily added to the model training. the reason we didn't include it because

  1. we found iou prediction is not very accurate on medical image segmentation (it could be that this branch is not well trained)
  2. we preferred to simplify the task by only including segmentation related modules.