search

binli123 / dsmil-wsi

DSMIL: Dual-stream multiple instance learning networks for tumor detection in Whole Slide Image
MIT License
332 stars 84 forks source link

Overlapping patches #68

Open jafarinia opened 1 year ago

jafarinia commented 1 year ago

Hi Thanks for the wonderful paper and very clear code. I have a question. Do you think that your model will still work if we give it overlapping patches? (I need to remove healthy patches from a bag of overlapping patches of WSIs on TCGA)

PiumiDS commented 1 year ago

Hi @hussein-jafarinia

Have you checked the image patches shared by Bin? It seems to me that the patches were already extracted with overlap as I saw the same patch repeated. I am confused about the patch extraction step so really appreciate your feedback.

Best regards.

jafarinia commented 1 year ago

Hi @hussein-jafarinia

Have you checked the image patches shared by Bin? It seems to me that the patches were already extracted with overlap as I saw the same patch repeated. I am confused about the patch extraction step so really appreciate your feedback.

Best regards.

By overlap I don't mean same patches repeating. I mean for example a patch having 50% overlap with other patches in it's horizontal and vertical axes (which is something that some methods on histopathology do at least for their training process) I haven't checked the details of their tiling/patching code for training(deepzoom_tiler.py). But the result of test_crop_single.py isn't overlapping (I've run it) and based on what the paper says and shows in it's figures it seems like their patches aren't overlapping.

Rainydu184 commented 1 year ago

it may be not work well in overlapping patches, the same area will influence the distance between the different patches.