search

jayrmh / QAERTS

We present QAERTS, a parameter-efficient multi-head model for 3D fetal brain pose estimation from freehand 2D ultrasound videos by leveraging uncertainty across different geometric transformations, for use in low-income settings.
6 stars 0 forks source link

Request for Data Access #1

Open cshennju opened 3 days ago

cshennju commented 3 days ago

I want to commend you on the fantastic work you've done with this project! It's incredibly insightful and helpful. I’m particularly interested in the trainning data (3D volume) you used and was wondering if it’s possible to share it?

jayrmh commented 3 days ago

Greetings @cshennju and hope you are well. Please note the following:

  1. You can download the publicly released final versions of our 3D atlases here for each week: https://www.nature.com/articles/s41586-023-06630-3. (format may be slightly different, but it's the same content).
  2. You no longer need a mask because the final atlas versions only contain the intra-cranial information! But if you are using our atlas, your volumes will need to be registered to a similar coordinate space. If they are not, consider this previous work: https://www.sciencedirect.com/science/article/pii/S1053811922004608.
  3. Using your own volumes, I suggest sampling a fixed set of 2D slices to place in the test folder. This is depending on how you would like to review the results. I looked at these specifically, i.e, the choroid plexus area, the lateral ventricles, the sylvian fissure, and images containing little fetal information to discard.
  4. If you just want to perform inference, just (1) is sufficient. If you need to retrain (2) is a must. (3) is optional based on your needs.
  5. The exact same training 3D volumes will have to be obtained from https://www.sciencedirect.com/science/article/pii/S0002937818300127 with the permission of the original authors.

Good luck, Jayroop

cshennju commented 3 days ago

Thank you so much for the detailed guidance and for sharing the links to the resources.

cshennju commented 3 days ago

Regarding the 3D training volumes from the paper 5 (https://www.sciencedirect.com/science/article/pii/S0002937818300127): Should I contact the INTERGROWTH-21st Consortium Executive Committee directly to request access to the data, or is there another process I should follow?