results not ideal

[TingTingShao]

Dear,

This is my first try with WSInfer, and I would like to ask for more features and advice on it.

The aim of the project:
(1) to identify metastasis sets, the statistics include: the number of the metastasis sets, and the area of each metastasis set.
(2) To automate the process, as it took so long with manual annotation.

I tried with SAM extension, it works great with Prompt, but can not perform good with Auto mask.

I tried with WSInfer annotation to see if it can help with the automation, but I can clearly see that the results are not ideal though it was fast. The metastasis sets were labeled with low probability of tumor. The model I used was breast-tumor-renet34.

The images are as follows:

Any idea on how to improve this?

Looking forward to your reply.
Thanks,
tingting

[kaczmarj]

what organ is the tissue from?

[TingTingShao]

Liver

[TingTingShao]

I can send one example image to you if that's better. BTW, is there a Discord channel for WSInfer?

Thanks,
tingting

[kaczmarj]

the main reason is that the tumor model you are using was trained on breast tissue. there's no guarantee that it would work well in liver, and i wouldn't expect it to work.

a solution would be to find a patch classification model for liver tumor. another option is to try qupath's built-in patch classification methods. if you opt to make your own liver tumor classifier, take a look at patch sorter https://github.com/choosehappy/PatchSorter

we don't have a discord server. most of the communication happens here in the github issues, sometimes in email.

[TingTingShao]

Many thanks for your reply.

Quick response to your first point:

the main reason is that the tumor model you are using was trained on breast tissue. there's no guarantee that it would work well in liver, and i wouldn't expect it to work.

This is the breast cancer metastate in liver. In this case, I should also use the model in Liver? Do you have a specific model in your mind for me to choose?

Thanks for other suggestions, I am gonna explore later.

Thanks,
tingting

[kaczmarj]

This is the breast cancer metastate in liver. In this case, I should also use the model in Liver? Do you have a specific model in your mind for me to choose?

ah i see. this is a good question. patch classification models can be finicky, so i think the best solution would be to have a patch classifier trained on breast metastases in liver tissue. from your screenshot, it looks like the breast tumor model is applying false positives in the hepatocytes.