How to segment bad quality 4D MRI FLOW

Support
1

Hi everyone :slight_smile:

I’m working on aortic segmentation from 4D flow MRI. I’ve been using TotalSegmentator but results are poor, I assume mostly a domain issue, since it’s CT-trained, plus the low SNR of my data.

Two things I’d love input on:

1. Anything off-the-shelf? A pretrained model or pipeline I could run directly on my volumes, without annotating or training anything. Has anyone tried the MR variant of TotalSegmentator, or a pretrained nnU-Net for the aorta, on 4D flow data?

2. If not, what’s the standard workflow? I’d rather follow an established approach than improvise. Do people segment the magnitude images directly, or reconstruct a PC-MRA first? How many annotated volumes are realistically needed to fine-tune something usable? Any public 4D flow dataset with aortic labels?

Happy to write up whatever ends up working, in case it’s useful to others.

Thanks a lot!