Resolution, magnification, and tiling#
Resolution choices determine both what a model can see and how much computation a workflow requires.
MPP and magnification#
MPP is a physical measurement; magnification is scanner and optics terminology. Common approximations are:
Nominal magnification |
Approximate MPP |
|---|---|
10× |
1.0 |
20× |
0.5 |
40× |
0.25 |
These are conventions, not a substitute for metadata. Inspect the slide before assuming a value:
print(wsi.properties)
print(wsi.fetch.pyramids())
If MPP is missing, provide the known source resolution through slide_mpp when tiling:
zs.pp.tile_tissues(wsi, 256, mpp=0.5, slide_mpp=0.25)
Do this only when the scanner or acquisition record provides a trustworthy value.
Field of view#
The physical width of a square tile is:
field of view in microns = tile_px * mpp
Tile |
MPP |
Field of view |
|---|---|---|
256 px |
0.5 |
128 µm |
512 px |
0.5 |
256 µm |
256 px |
0.25 |
64 µm |
Changing tile pixels and changing MPP are therefore not equivalent. Choose the input expected by the model whenever using a pretrained model.
Stride and overlap#
With no stride_px or overlap, adjacent tiles do not overlap. Smaller strides improve spatial coverage but increase tile count and compute.
# 25% overlap
zs.pp.tile_tissues(wsi, 256, mpp=0.5, overlap=0.25)
# Equivalent explicit stride for square 256-pixel tiles
zs.pp.tile_tissues(wsi, 256, mpp=0.5, stride_px=192)
overlap and stride_px are alternatives; specify one, not both.
Pyramid levels#
Use level for operations that work directly on a stored pyramid level, such as traditional tissue detection. Use MPP for algorithms whose behavior must be physically comparable across scanners.