Open shntnu opened 5 months ago
Hi @shntnu, was this information published anywhere yet? Thank you!
I looked into metadata,
for sources 1,3,4,11 there is Index.idx.xml
, tags ImageResolutionX
, ImageResolutionY
for source 9 indexfile.txt
table file ImageSizeX
, ImageSizeY
columns
for sources 2,5,6,10,13 MeasurementDetail.mrf
, tags HorizontalPixelDimension
, VerticallPixelDimension
.
Unfortunately I did not find metadata files for source 7 and 8.
(source 7 and 8 updated by @shntnu)
source | μm |
---|---|
source_1 | 0.597976 μm |
source_2 | 0.65 μm |
source_3 | 0.597976 μm |
source_4 | 0.597976 μm |
source_5 | 0.64974 μm |
source_6 | 0.646187 μm |
source_7 | 0.65 μm |
source_8 | 0.7032 μm |
source_9 | 0.597976 μm |
source_10 | 0.6516944 μm |
source_11 | 0.597976 μm |
source_13 | 0.65 μm |
Thanks for documenting that @Arkkienkeli! I have updated your table
source_13
and source_7
were from the same data generating center, so they are identical.source_8
is 0.7032 μm (see below)I'll leave this issue so we one of us comment on what the implications are (briefly – one of the reasons normalization is important)
To double-check my understanding, this means that we need to resize the images from different sources according to the table?
@ChenyuWang-Monica nope, definitely not! Will expand below.
@shntnu and @Arkkienkeli , thanks for digging in on this! Can we add this to the metadata CSV?
For normalized measurements, this will have almost-no impact, since most things are in pixel units, and fundamentally, once we're normalized, they're in "mean cell units". A ~16% change in the pixel size (largest/smallest ~= 1.16) will maybe change some precision, but not much else.
For non-normalized measurements, realistically we shouldn't be comparing across sites anyway, but even if we are, for things like area, it would just be applying the pixel conversion factor before comparing. The only place where things get a littttle dicey is stuff around textures and granularity, but we're typically looking at scales around 3/5/10 pixels, so a 1.16 multiplier still only changes things one or two pixels, give or take (3->3.48, 5->5.8, 10->11.6). Realistically, I'd expect other experimental details for those metrics, since they're intensity dependent, would dominate over the pixel size discontinuity.
@bethac07 Thanks for the explanation! I have two follow-up questions:
For non-normalized measurements like area, is the adjustment already applied to the well-level CellProfiler features in workspace/profiles? If not, is there a file describing whether each feature is normalized or non-normalized?
If I want to use the raw images instead of the features, how should I adjust for such resolution differences? Would resizing the images according to the resolution work?
All measurements start off non-normalized and then get aggregated, then normalized, then feature selected. It should be pretty clear in the file name which is which. All measurements are in pixel units, nothing we report has a pixel-to-micron conversion applied.
The interpolations involved in resizing images I suspect will be more destructive than the pixel size difference, so I would personally suggest you not do that - again, it's a relatively small difference (16%). I wouldn't try to correct for it but more just note that it might be the source of small pixel size discrepencies when looking across sources.
CW asked: