Bad Segmentation with Large IHC Images

[Salmck]

I want to perform cell segmentation on a 3000x3000 IHC image as input, with a tile_size of 300. The big one will be cropped into 100 smaller images, but some of them containing cells are mistakenly identified as background. How should I perform the accurate segmentation?
these are some examples:

[sanadeem]

@Salmck can you please share the original 3000x3000 images and we can give it a try. We haven't seen this before. Did you try running on our https://deepliif.org by uploading the image (selecting 5X resolution)?

[Salmck]

@Salmck can you please share the original 3000x3000 images and we can give it a try. We haven't seen this before. Did you try running on our https://deepliif.org by uploading the image (selecting 5X resolution)?

@sanadeem These are some images from my IHC dataset from hospital. I have tried on your website before. Nearly half of the SegOverlaid images are incomplete.

[jommarin]

@Salmck The image quality seems to be too low in these images for our model to handle as-is. You can try pre-processing the images to improve the contrast and running it on our website at 5x (tile size of 128 if you are running the code). I used the auto-contrast adjustment in Photoshop and that seemed to help.

[Salmck]

@jommarin okay, I will try later.