point prompt performance vs original SAM

[sedghi]

I'm running the tutorial_point_prompt_seg.ipynb notebook and experimenting with different slices, but the results are worse than the original SAM. I would appreciate any insights into this issue.

MedSAM

Slice 85

https://github.com/bowang-lab/MedSAM/assets/7490180/92096589-0dec-405b-9e0a-c05dd5482108

Original SAM (Meta)

https://github.com/bowang-lab/MedSAM/assets/7490180/59817253-051d-4828-a767-b104504ab9d3

[MaryaEb]

hello,

Could you please confirm if you found the notebook in the SAM/MEDSAM repository, or if you're working on it separately? I'm currently focusing on refining the point prompt using the MEDSAM model.

[sedghi]

Yeah I used the notebook in the repo, and only changed the slice number from the middle slice to another slice

[JunMa11]

Hi @sedghi ,

This is expected since the point prompt is still an experimental function and the model was trained on a small abdomen CT organ segmentation dataset. In fact, this is still an open research question. The goal of sharing the training script is to prompt the development in this direction.

[sedghi]

It's good to know. I assumed that since vanilla SAM works fine on the CT, MEDSAM would be even better

[MaryaEb]

@sedghi, I couldn't locate the tutorial_point_prompt_seg.ipynb notebook. Could you please share it with me?

[sedghi]

it is here https://github.com/bowang-lab/MedSAM/tree/main/extensions/point_prompt

[MaryaEb]

Thank you @sedghi