Lightsheet_microscopy_image_3D_nuclei_instance_segmentation
Lightsheet microscopy whole brain 3D nuclei instance segmentation
2D Unet-based
Motivations:
- None has developed a high-throughput nuclei instance segmentation for whole brain.
- Lightsheet microscopy 3D image has anisotropic resolution.
- Resolution is isotropic in X-Y plane, where human annotating nuclei, then tracking Z stack.
Performance and time cost:
- Recall > 90%
- Precision > 90%
- Time cost ~= 15hr/brain
Data:
- We segmented on multiple whole brains of mouse in different grown stage, P4 and P14.
- P4 brains have ~1500x9000x9000 voxels, and about 30,000,000 to 50,000,000 nuclei.
Example to test one partial brain:
1 Obtain NIS results of brain
cd cpp, then follow the instruction in cpp/README.md.
2 Postprocess NIS cpp output
cd ..
python coord_to_bbox.py --nis_output_root downloads/cpp_output
This will generate bounding box of NIS in the output folders for stitching.
usage: coord_to_bbox.py [-h] [--device DEVICE]
[--nis_output_root NIS_OUTPUT_ROOT]
Generate NIS bounding box
optional arguments:
-h, --help show this help message and exit
--device DEVICE
--nis_output_root NIS_OUTPUT_ROOT
Output dir of NIS cpp
3 Image stitch
Check paths in image_stitch/stitch_main.py, then
python image_stitch/stitch_main.py
3.5 (Optional) Refine stitch by point registration
Check paths in image_stitch/ptreg_stitch_p14.py, then
python image_stitch/ptreg_stitch_p14.py
4 Generate brain maps
python gen_brain_map.py will save a brain map as Nifti under downloads/cpp_output/test_pair
5 (Optional) Colocalization with additional channels
See coloc/README.md
Web app: Interactive visualization of whole brain nuclei segmentation results
- Check on frontend repo
Codes availability
- Train 2D Unet
- Source code of the executable to test a whole brain
Data availability
See https://bossdb.org/project/curtin2026