Open OCEANOUXIN opened 5 months ago
Hi @OCEANOUXIN
Good question. In my opinion, synthetic tumors in this paper are fundamentally different from data augmentation.
Suppose you have 100 CT volumes with liver tumors from the LiTS dataset. With data augmentation techniques like rotation, blurring, flipping, and adding noise, you can make minimal changes to these 100 volumes—same tumors, same patients.
However, tumor synthesis can be applied not only to these 100 volumes but also to numerous healthy CT volumes, which are more common in clinical practice. This approach can generate thousands of CT volumes with per-voxel annotations—different tumors, different patients.
Therefore, tumor synthesis is a promising research direction.
Best, Zongwei
Hi @OCEANOUXIN
Good question. In my opinion, synthetic tumors in this paper are fundamentally different from data augmentation.
Suppose you have 100 CT volumes with liver tumors from the LiTS dataset. With data augmentation techniques like rotation, blurring, flipping, and adding noise, you can make minimal changes to these 100 volumes—same tumors, same patients.
However, tumor synthesis can be applied not only to these 100 volumes but also to numerous healthy CT volumes, which are more common in clinical practice. This approach can generate thousands of CT volumes with per-voxel annotations—different tumors, different patients.
Therefore, tumor synthesis is a promising research direction.
Best, Zongwei
Thanks for your reply! I also wonder if the TumorGenerated.py could generate the synthetic liver label without the GT? I see the code seems to have to use the GT. That's means it's a method of supervision.
The GT you referred to was randomly generated shapes. These shapes are used to control the generation of tumors and are not annotated by humans.
The GT you referred to was randomly generated shapes. These shapes are used to control the generation of tumors and are not annotated by humans.
I still confused, how to get the synthetic liver? Isn't it from the TumorGenerated.py? But in your transform function you get the GT before the use the TumorGenerated.py. Could you explain it? Thank you so much!
Pardon?
@OCEANOUXIN The GT we use is organ segmentation, which is almost free for the liver. The code is for generating liver tumors—two different things.
@OCEANOUXIN The GT we use is organ segmentation, which is almost free for the liver. The code is for generating liver tumors—two different things.
Thank you for your reply! I understand now.
Thanks you for your excellent works, I want to know whether this work is a equal way to augment data.