search

MrGiovanni / DiffTumor

[CVPR 2024] Generalizable Tumor Synthesis - Realistic Synthetic Tumors in Liver, Pancreas, and Kidney
https://www.cs.jhu.edu/~alanlab/Pubs24/chen2024towards.pdf
Other
134 stars 10 forks source link

The use of organ_pseudo #1

Closed Luffy03 closed 7 months ago

Luffy03 commented 7 months ago

Hi, congratulations on your excellent work first! It inspires me a lot! I want to know what is the use of 'organ_pseudo' in validation. It seems that you are using the organ_pseudo to filter out the tumor predictions, is it correct? And how to generate organ_pseudo? Train another organ segmentation model to predict it or directly draw from the real labels?

qic999 commented 7 months ago

@Luffy03 Thanks for your attention to our work!

Post-processing operations are commonly used in tumor segmentation. Due to the similarities of tumors across different organs to some extent, it is necessary to filter out tumor predictions in non-target organs. In our study, we use the public pre-trained organ segmentation model from https://github.com/ljwztc/CLIP-Driven-Universal-Model to generate organ pseudo labels. We then use these pseudo labels to ensure that tumor predictions correspond to the correct organ.

It is acceptable to use other pre-trained organ segmentation models as well. Our experiments are fair since all results are obtained using the same post-processing operations.

Luffy03 commented 7 months ago

Ok, thanks for your answer.