Segmentation Issues with Ovarian Pathology Images

[anamika-yadav99]

Dear Authors,

Thank you for releasing this model as open-source! It's a valuable resource for the community.

I am attempting to segment non-pathological (normal) ovarian pathology image (H&E stain) using the provided example_prediction.py script. My WSIs have a magnification of 20x and an MPP (Micrometers Per Pixel) of 0.4942 for both X and Y axes. I've created image patches of size 1024x1024 pixels for segmentation. My goal is to segment epithelial cells, connective tissue cells, neoplastic cells, stromal cells, and inflammatory cells.

However, the current segmentation results fail to segment the cell type of interest. I've tried adjusting the base magnification within the script from 20x to 40x, but this hasn't improved the segmentation.

Would you be able to offer any suggestions on the best parameters to use for segmenting ovarian pathology images with this model? I've attached a few example slides for your reference.

[theodore-zhao]

For the pathology slides, I would recommend cutting the images into even smaller patches for best performance, as that was how the model was pretrained. You can try cropping to 256x256 patches, and resize each patch to 1024x1024 for model inference. Hope that can help the performance!