A user would prefer to extract tiles at a specified resolution in micrometers per pixel rather than at the openslide level. Available levels (0,1,2) can mean very different resolutions between scanner types and even scanner settings.
Our answer below:
This is more tricky than you may realize. our other work using .e.g, histoqc (http://histoqc.com/), has demonstrated that the same scanner can produce different MPP at the same magnification for different slides in some cases these differences are ~5-10% difference, e.g., .2456 vs .2468 vs .2325 in the TCGA alone. Thus the question arises, if you specify a MPP, do you want a .2456 to be scaled to a requested e.g., .23? I suspect this will introduce a large number of interpolation type artifacts, and would actually negative impact DL robustness. proper study is needed in this context to determine what the best-practice should in fact be. i opted to go with the openslide level because it is both (a) more transparent, you get exactly the raw data from the file and (b) it is harder to do, hopefully requiring people to think about these questions instead of blindly forging ahead that said, our colleague has produced a WSI library which will allow you to do this and should be fairly easy to integrate into the provided code: https://github.com/pjl54/WSI_handling
A user would prefer to extract tiles at a specified resolution in micrometers per pixel rather than at the openslide level. Available levels (0,1,2) can mean very different resolutions between scanner types and even scanner settings.
Our answer below: This is more tricky than you may realize. our other work using .e.g, histoqc (http://histoqc.com/), has demonstrated that the same scanner can produce different MPP at the same magnification for different slides in some cases these differences are ~5-10% difference, e.g., .2456 vs .2468 vs .2325 in the TCGA alone. Thus the question arises, if you specify a MPP, do you want a .2456 to be scaled to a requested e.g., .23? I suspect this will introduce a large number of interpolation type artifacts, and would actually negative impact DL robustness. proper study is needed in this context to determine what the best-practice should in fact be. i opted to go with the openslide level because it is both (a) more transparent, you get exactly the raw data from the file and (b) it is harder to do, hopefully requiring people to think about these questions instead of blindly forging ahead that said, our colleague has produced a WSI library which will allow you to do this and should be fairly easy to integrate into the provided code: https://github.com/pjl54/WSI_handling