Nuset Segmentation¶
Deep learning based segmentation, useful for nuclear localized indicators. ROIs segmented through this module can be imported into the Viewer Work Environment.
Note
If you use this tool, please cite the Nuset paper in addition to citing Mesmerize: Yang L, Ghosh RP, Franklin JM, Chen S, You C, Narayan RR, et al. (2020) NuSeT: A deep learning tool for reliably separating and analyzing crowded cells. PLoS Comput Biol 16(9): e1008193. https://doi.org/10.1371
Parameters¶
Projection¶
Choose a projection which maximizes the visibility of your regions of interest
Pre-process¶
Parameter |
Description |
|---|---|
do_preprocess |
perform pre-processing |
do_sigmoid |
perform sigmoid correction |
sigmoid_cutoff |
cutoff, lower values will increase the exposure |
sigmoid_gain |
gain, high values can be thought of as increasing contrast |
sigmoid_invert |
invert the image if necessary. Regions of interesting should be bright, background should be dark |
do_equalize |
perform adaptive histogram equalization |
equalize_lower |
Set a lower limit, this helps remove background & increase contrast |
equalize_upper |
Upper limit for the histogram |
equalize_kernel |
kernel size, increase if the pre-processed image is grainy. Start with a value ~1/16-1/8 the size of the image |
NuSeT¶
Parameter |
Description |
|---|---|
watershed |
wastershed the image, useful if your cells are tightly packed. Uncheck if cells are large and/or sparse. |
min_score |
Decreasing this value will cause more regions to be found, i.e. cells tend to split more |
nms_threshold |
Increasing this value will cause more regions to be found, i.e. cells tend to split more |
rescale_ratio |
Use smaller values less than 1.0 if you have large bright cells, If you have smaller or dim cells use values higher than 1.0 |
Note
min_score & nms_threshold work in opposing ways
Note
Segmentation will utilize all threads available on your system (regardless of the value set in your System Configuration). However it only takes a few seconds or a few minutes if segmenting a large 3D stack.
Note
high rescale_ratio values will increase the time required for segmentation. Values around 3.0 take about ~1 minute for 512x512 sized images on ~16 core CPUs.
Post-process¶
Export¶
If you export using a Convex Hull masks containing only a few pixels, which may be noise, will be removed.
Note
Segmentation will utilize all threads available on your system (regardless of the value set in your System Configuration). However it only takes a few seconds if exporting a 2D image, and make take ~10 minutes if exporting a large 3D stack.