search

arcadelab / FastSAM3D_slicer

A 3D Slicer extension for FastSAM3D
https://arxiv.org/abs/2403.09827
11 stars 2 forks source link

About Segmentation effect #3

Closed jianjun0407 closed 3 weeks ago

jianjun0407 commented 3 months ago

I directly use your model in 3D slicer directly with prompts, I found it really fast, almost in real time. But, on the other hand, I found that the segmentation was not very good, and I recorded two videos, one trying to segment the lung and one trying to segment the kidney. Want to ask, is there any problem with my use, causing bad segmentation effect ?

https://github.com/arcadelab/FastSAM3D_slicer/assets/19251556/cc0426ca-d9bd-4717-9560-d6b04ee93a25

https://github.com/arcadelab/FastSAM3D_slicer/assets/19251556/3fc0a5e5-f1d1-4d67-b247-dfce028ed17e

skill-diver commented 3 months ago

You may need finetune in your own dataset to get better performance. 3d volume data is hard to segment so you may take lots of time to do finetune.

jianjun0407 commented 3 months ago

Yes, I agree with you. For the segmentation of lung, it seems that there is no in the training data set. However, the segmentation categories mentioned in the experiment of the paper, such as kidney segmentation, which is available in the TotalSegmentator dataset, also show poor results from the second video mentioned above. I would like to ask, what do you think might be the reason?

skill-diver commented 3 months ago

Do you resmaple the data? If you want better performance you need to use the resample script to change resolution of your image. The model don't have comparely good effect on too high resolution image.

skill-diver commented 3 months ago

https://github.com/arcadelab/FastSAM3D/blob/main/utils%2Fprepare_uunet.py

Here's the resample script.