search

bowang-lab / MedSAM

Segment Anything in Medical Images
https://www.nature.com/articles/s41467-024-44824-z
Apache License 2.0
3k stars 416 forks source link

about dataset ensemble #244

Closed PANHAOC closed 7 months ago

PANHAOC commented 7 months ago

Hello, I would like to know if the labels for using these datasets are consistent? For example, for the task of segmenting lung lobes, in dataset A, its label ID is 1, while in dataset B, its ID is 3. Does the difference between these two IDs have an impact on the training process? If there is an impact, how do you handle unifying it?

JunMa11 commented 7 months ago

The direct answer is no.

You can directly merge different datasets without changing the label id.

PANHAOC commented 7 months ago

The direct answer is no.

You can directly merge different datasets without changing the label id.

Thank you for your reply. I also want to ask, how is the training process for a dataset with multiple labels? Is each label treated as the foreground and the rest as the background separately, and are slices made for each label separately？

JunMa11 commented 7 months ago
  1. how is the training process for a dataset with multiple labels? Is each label treated as the foreground and the rest as the background separately

Not exactly. MedSAM is box promptable method, the label is associated the given box prompts.

  1. are slices made for each label separately？

No. the data loader will automatically sample labels from the multi-label mask.