[FEATURE] trouble detecting the "brightest" cells in image

[EdwardNeuro]

Hi, I'm using cellpose 3.1.1.2 to detect my cells but got some trouble. After training with my own 26 images and masks, it successfully improved precision, but couldn't detect the "brightest" cells, as is shown in attached screenshots, including two auto and one mannually detected.

I've changed my training parameters like learning rate and epochs, but got similar results. I'm wondering if anyone confronted similar problems and which steps I could make mistakes.

Thanks a million!

[mrariden]

@EdwardNeuro I'm not sure, you can try to use CP4. Overall, the performance is better.

[EdwardNeuro]

Thank you so much, I've tried big data (more than 20GB) using cellpose.contrib.distributed on SAM-Cellpose, but it always caused out of GPU memory (NVIDIA RTX 5880 Ada 48G). That's why I turned to cp3😂and successfully run my code. It seems like cp3 can automatically adjust GPU usage, but SAMCP4 should be set somewhere manually? @mrariden

[mrariden]

If you have very large images the normalization may be thrown off which could be related to your issue. See #1277

Cellpose needs to make multiple copies of the image and holds multiple copies of intermediate representations all of which use memory. CP3 has a different architecture and a smaller model which is why you can use it on your data. I'd recommend experimenting with smaller images/crops/smaller data types.