[BUG] Square-shaped artifacts in Cellpose flow

[thewoz]

Before you fill out this form:

Did you review the FAQ?

• Yes, I reviewed the FAQ on the cellpose ReadTheDocs

Did you look through previous (open AND closed) issues posted on GH?

• Yes, I searched for related previous GH issues.

Hi everyone,
I’m using Cellpose 4.0.8 and I’m encountering an issue I can’t figure out. I’m working with a very large image (~14,000 × 14,000) and running cell detection on it.
In the flow output, I observe clear square-shaped artifacts that are causing problems downstream. I’ve attached a close-up showing the issue. As you can see, the cells themselves are detected correctly.
I’m not sure what is causing these artifacts or how to eliminate them. Any insight or suggestions would be greatly appreciated.

[mrariden]

The network in cellpose 'sees' images in 256² tiles, regardless of the actual image size. So, this intermediate output is somewhat expected. The flows are scaled by the cellprobability prediction, so the final cell masks shouldn't have low-likelihood masks.

You can also look at the blocksize parameter, which does tile-wise normalization.

[thewoz]

Thanks for your reply.
If I understand correctly, the image is divided into 256×256 tiles, and the tile size can be adjusted using the blocksize parameter.
Are the flow values scaled by cellprobability before or after being displayed?
What I don’t understand is how very large, clearly incorrect, flow vectors can be detected in the background of one tile but not in another tile.

[nathantspencer]

I am encountering a similar issue in a smaller image, in which the intensities of background cell probabilities and flows vary significantly from tile to tile in ways that aren't obviously discernible from the original image.

The panels below are the original image at top-left, resulting masks at top-right, flows at bottom-left, and cell probabilities at bottom-right.

Some of these appear to result in nearly tile-sized masks. I can filter the unwanted results out by size, but given the unusual look of the intermediate images, I'm hoping to better understand whether this is an expected artifact that should be handled post-prediction vs. something I should be looking to address earlier in the process.