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 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.
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.