search

ljwztc / CLIP-Driven-Universal-Model

[ICCV 2023] CLIP-Driven Universal Model; Rank first in MSD Competition.
Other
581 stars 71 forks source link

provided post label shape mismatch #82

Open skapoor2024 opened 3 months ago

skapoor2024 commented 3 months ago

Hi I was trying to recreate your experiment. I found out there are multiple labels of pancreas in post label which are not available on TCIA_pancreas provided segmentation mask. It just contains those of pancreas. Can we get those multiple segmentation masks. Also the provided post label shapes are (31,d,h,w) meanwhile there are 32 organs as mentioned in the table. This issue is there with every provided dataset label apart from 12_CT-ORG and 05_KiTS. Moreover in KiTS maybe the voxel size of cyst is small and the post label preprocessing makes them zero for the 32th label.

Also why in post labels the organ that are not present all the values in those slices are 255.

ljwztc commented 3 months ago

Each channel in postlabel indicates the binary mask for corresponding label. We initially trained the model with 31 organs, and later added one more organ (32 Kidney Cyst). And we allocated one data in one card, so it was smooth when we run the experiment. Indeed, if we run two data in one card, there would exist the mismatched shape problem. So 32 organs post-label should be regenerated for whole dataset. Or you can simply run 31 organs.