search

OpenGVLab / SAM-Med2D

Official implementation of SAM-Med2D
Apache License 2.0
792 stars 75 forks source link

Question regarding the dataset at HuggingFace #63

Open anwai98 opened 1 month ago

anwai98 commented 1 month ago

Hi team,

Thanks for open-sourcing your amazing effort.

I have a question: I managed to download the dataset you provided hosted at https://huggingface.co/datasets/OpenGVLab/SA-Med2D-20M, and looks like only a part of the dataset has been released (composed of ~3.7M images and ~15.8M masks).

Do you plan to provide access to the rest of the images (according to the paper, it is marked as the "test set", comprising ~0.92M images and ~3.9M masks)? Would be nice to check it out as well.

Thanks in advance!

anwai98 commented 1 month ago

Hi team,

Would like to share a few mentions and verify if this is the case or not:

  1. Some input images from the Brain_PTM dataset appear a bit weird (for example: images/mr_t1--Brain_PTM--case0005--x_0052.png, images/mr_t1--Brain_PTM--case0005--x_0056.png, a few of the many which appear to be "binary mask of the brain" as the "input image")
  2. Some images in QUBI2020 dataset have a similar strange appearance, especially the "brain_growth" samples (for example: images/ct_00--QUBIQ2020--1_brain-growth_case01--2d_none.png, images/ct_00--QUBIQ2020--1_brain-growth_case25--2d_none.png, a few of the many which do not appear as the tissue region itself, rather a binary-ish visual)
  3. ~For the autoPET dataset, the images are formed using the CT scans-only right?~
    • EDIT: Missed this one, apologies. The PET scans are available as a separate image paired with the lesions under the modality pet.
  4. Some images have mismatching shapes for their respective ground-truth (I could spot only two at the moment): x_ray--covid_19_ct_cxr--auntminnie-2020_01_31_20_24_2322_2020_01_31_x-ray_coronavirus_US--2d_none.png and x_ray--covid_19_ct_cxr--radiopaedia-2019-novel-coronavirus-infected-pneumonia--2d_none.png

I'll come around with more questions, if any.

Thanks!