search

HiLab-git / WORD

[MedIA2022]WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from CT image
https://www.sciencedirect.com/science/article/pii/S1361841522002705
GNU General Public License v3.0
143 stars 19 forks source link

can't reproduce the segmentation results of left femoral head and right femoral head #4

Closed Ctozzzzz closed 11 months ago

Ctozzzzz commented 12 months ago
I've tried training both 2D nnUNetv2 and 3D nnUNetv2 using the WORD dataset. But I found that the segmentation results of Head of femur were not as good as illustrated in the paper. The result of 3D nnUNetv2 were: target DSC HD95
Head of femur(L) 0.545±0.114 196.485±22.436
Head of femur(R) 0.471±0.138 173.113±35.954

If you have met this situation, I wonder if you can offer me an explanation or guideline. Thank you😀

Luoxd1996 commented 12 months ago

Hi, It is caused by the flipping augmentation, there are two solutions: turn off the flipping augmentation or post-processing. Best, Xiangde Luo.

Ctozzzzz commented 12 months ago

Hi, It is caused by the flipping augmentation, there are two solutions: turn off the flipping augmentation or post-processing. Best, Xiangde Luo.

I will try it. Thank you so much!