hi, after downloading the slides I opened each of them in order to retrieve slide spacing at level 0 (in micron per pixel – mpp) ; however, all slides have unusual spacing values: 1000
judging from a few slides I visually looked at (comparing with slides from a different dataset where spacing is known), it seems the slides you provide have 0.25 mpp at level 0
I was expecting different spacing based on the center the slides comes from, but it seems all slides have the same same spacing: do you confirm all slides have 0.25 mpp at level 0?
Hi,
so, the spacing value of 1000 in the slides has no relation to the actual pixel spacing? If this is the case, is there a way to get the true spacing at level 0 for a slide?
Thank you for your prompt response. However, it is crucial to know the actual spacing per slide, as it can greatly limit the strategies for developing an accurate prediction model. To maximize the potential of this challenge and achieve higher performance, it would be advantageous to have the actual spacings per slide available.
One possible solution could be to provide a CSV file with the spacings per level for each slide in the training and test set. This would be a fast and effective way to provide this information to participants. Nevertheless, it is important to emphasize that the spacing information should also be readable from the files in the hidden evaluation set. Providing this information can ultimately lead to improved results in the challenge.
hi @ishashah, thank you for your answer.
is there a way one can access the actual spacing of slides at level 0?
not knowing the spacing is blocking me from going further in the challenge given convolutions are not scale invariant (hence if I want to train a neural network that has convolutional layers, I’ll need to provide images at a fixed scale – that is at a fixed spacing). as @daangeijs suggested you could maybe provide an additional file with the true spacing at level 0 for each slide.
Thanks for bringing this question to us, and for explaining what would be helpful - we have been discussing this with our partners and will have an answer for you very soon.
I am also following the issue and in fact the images with mpp=0.25 and 0.5, at level 0, have significant difference.If you can provide that is very good
Hi @ishashah,
Thanks for the prompt action and the added column of resolutions. This solves a lot for now, however can you confirm that the files of the evaluation set will contain the right spacing, in other words if we upload our algorithm and start processing the files will we be able to read the correct spacing from those those files?
The evaluation set will have a metadata file associated with it (test_metadata.csv) that will also have a resolution column, so your solution will be able to read a resolution value from that file. To see an example, you can look at test_metadata.csv in smoke test sample data code_execution_development_data.tgz on the Data Download page.
Is the length-per-pixel ratio constant among the same page/quality?
In other words, for each page/quality q, is there a constant c[q] such that the length-per-pixel at quality q = c[q]*test_metadata[‘resolution’]?
I am asking because there is a relatively large gap between the size of the images at page/quality 4 for example: in jpeg format, it’s from 27mb the largest to 100 kb the smallest image.
The images do indeed vary greatly in size - if you look at the raw formats, you’ll see that the range from < 50 Mb to more than 4 GB, and this would be reflected in the size of images at page = 4.
Each page of the pyramidal tif is half the size of the previous one, so if you take the resolution at page = 0 (which is what we provide in test_metadata["resolution"]), it will be 4 times sharper than the resolution at page = 1. This relationship also holds between every other sequential pair of pages.