Question about kidney subregion cropping in KiTS23 challenge winner solution

[uturkbey-ryverai]

Hi,

I’m currently trying to reproduce the 1st place solution in the KiTS23 challenge based on Auto3DSeg for a research project.

I have already checked the paper by the NVAUTO team and the 'Differences with 1st place KiTS23 solution' section in the Auto3DSeg KiTS23 tutorial, but could not find the answers to the following questions:

1. When training the second-stage subregion prediction models with "Kidney subregion" cropped CTs, did they crop the training images using ground-truth bounding boxes, or did they use the bounding-box predictions from the first-stage foreground prediction models?
2. Did they handle the validation images any different from training images when cropping for the second-stage?
3. For the first-stage, did they train only SegResNet models, or did they also use DiNTS in an ensemble as in second-stage?

I’d really appreciate any help or guidance you can provide. Thank you so much!