Interest in training a marmoset fMRI model.

[Tyler Cook]

Hello Dr. Yeh,

I have used U-Net studio to perform some segmentation on marmoset T2w images, which worked fantastically. I am very interested in expanding this for use on functional images and training a model capable of performing extraction on fMRI EPI images. I am wondering if you have any advice on what input data would be suitable for this endeavour. Please let me know, thank you.

[Frank Yeh]

Hi Tyler,

    Sorry for the late reply. The input data would be a set of averaged fMRI EPI image. I will see if I can help train a model for this purpose.

Best,

Frank

On Thu, Jul 31, 2025 at 11:30 AM Tyler Cook <tyleral...@gmail.com> wrote:

Hello Dr. Yeh,

I have used U-Net studio to perform some segmentation on marmoset T2w images, which worked fantastically. I am very interested in expanding this for use on functional images and training a model capable of performing extraction on fMRI EPI images. I am wondering if you have any advice on what input data would be suitable for this endeavour. Please let me know, thank you.

--
You received this message because you are subscribed to the Google Groups "UNet Studio" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unet-studio...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/unet-studio/c9c28d51-563f-4ce0-a85d-8412dc83789fn%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Frank Yeh]

I can try using averaged volume within one subject to see if it works.

Do you know any public dataset available?

Frank

On Fri, Aug 8, 2025 at 12:30 PM Tyler Cook <tyleral...@gmail.com> wrote:

Hi Frank,

Thanks for the response. A model for the EPI images would be fantastic; the reduced resolution often makes it difficult. It would save our lab an enormous amount of time. I'm happy to help in any way if needed.  When you say averaged, do you mean multiple volumes within one subject, or multiple subjects & multiple volumes formed into a mean image? Thanks again.

[Frank Yeh]

I tested the marmoset.t2w.seg5 model, and it seems to work on bold data (see below).
I extracted the first volume from the BOLD series (current version does not  take 4D, but I can revise it) and used probability threshold of 0.3 to get more larger coverage:

Frank