MedSAM seems to be poor at segmenting bilayer membrane structures

[HaoyuCao]

Dear Authors:

I am a researcher of deep learning algorithms and level set algorithms. Currently, I have developed a set of image segmentation models for bilayer membrane structure based on dual level set approach. Comparative experiments on SAM and MedSAM models are performed on ACDC MMs and Refuge2 datasets.

The datasets are labelled as nested bilayer circular structures(i.e. 0:background, 1:outer membranes, 2:inner membranes). We obtain the best performance of SAM and MedSAM by manually specifying the bounding box for inner and outer membranes separately, and we subtract the small box prediction from the large box prediction to obtain the final inner and outer membrane prediction. However, in our experiments we found that MedSAM's segmentation of bilayer membrane structures is much weaker than the B/L/H version of the SAM model (the difference is about 20% in Dice Score).

I would like to ask whether MedSAM has some shortcomings for segmenting bilayer membrane types and whether you have considered further optimisation or fine-tuning of the publicly available weighting model.

[JunMa11]

Hi @HaoyuCao ,

Bounding box prompts are not a good choice for myocardium-like structure. Scribble prompt is a better choice for this task.
https://github.com/bowang-lab/MedSAM/tree/LiteMedSAMScribble