Question about the num of patches with mitosis in MIODG daataset

[Shentl]

Hi, @Xiyue-Wang

In the paper, you introduce the MIDOG dataset as "The publicly available training set contains 150 WSIs with a size of 8,000 × 8,000 pixels, which are cropped into 79,399 patches (6,699 with mitosis) with a size of 256 × 256 pixels" and I have some queestions

Assume you crop WSI based on the grid:

1. There are a total of 7149 annotations, of which only 1721 are for the "mitotic figure". So how did you obtain 6699 mitotic patches, as most of the "mitotic figure" bounding boxes only fall within one patch

2. Assuming you crop WSI based on the grid (for i in range(0，shape[0]/256): ......) instead of randomly cropping, the resulting number of patches should be 79426, which is 27 more than what was mentioned in the paper. And i wonder what happens

[Shentl]

Sorry , the total number of annotations is 4135, and only 1721 annotations are for the "mitotic figure"

[Xiyue-Wang]

@Shentl please download MIDOG dataset , Not every image size is 8000x8000, but the average size is 8000x8000

[Shentl]

I konw it, and 79426 is calculated using the actual image size

[Shentl]

What I would rather know is how 6399 mitotic patches were obtained as only 1721 annotations are for the "mitotic figure"

[Xiyue-Wang]

‘So how did you obtain 6699 mitotic patches, as most of the "mitotic figure" bounding boxes only fall within one patch’：Crop the patch overlap 50%, and then there will be some boundary processing

[Shentl]

By croping the patch overlap 50%, the total number of patches would be approximately twice of 79426, so I guess this is only done for areas labeled with mitosis

[Xiyue-Wang]

Right

[Shentl]

By "Crop the patch overlap 50%", at the inference time, a mitotic area is very likely to have multi bounding boxs. May I ask how you handled it? Did you directly use all the boxes to calculate the F1 score, or did you merge different bounding boxes?

[Shentl]

Can you tell me how you do at the inference time? If still crop patches overlap 50%, then how to deal with the multi bounding boxs for just one mitotic area?