3d_nuclei_seg_benchmark

Benchmarking of segmentation methods for 3D images of plant nuclei. Images are segmented. The segmentation is then evaluated with metrics (see 'metrics' section for more details).

dataset

3D images of plant nuclei. An image contains a single nucleus. The ground truth segmentation was partially done thanks to NucleusJ and manually.

metrics

• 3D jaccard index.
• 3D dice index.
• Nucleusj morphology metrics.

setups

Each folder has to be setup with a Docker specific environment.

todo

2d (applied slice by slice)

1. [2d_maskrcnn] Mask-RCNN: https://github.com/matterport/Mask_RCNN
2. [2d_topcoders] DSB2018 winner team (3 codes available: selim, albu and victor): https://github.com/selimsef/dsb2018_topcoders
3. [2d_stardist] Stardist: https://github.com/stardist/stardist

3d

1. [3d_stardist] Stardist (no pretrained models...): https://github.com/stardist/stardist

work in progress

2d

1. [2d_maskrcnn] trying...
2. [2d_topcoders] DSB2018 winner team: selim's code outputs only black images (still work in progress)

results

For now the results are computed on the OMERO_FSU dataset with the Otsu segmentation considered as the ground truth. This will be change later.

Method	2D or 3D?	Framework	Avg. Jaccard	Avg. Dice	Avg. F1	NJ2	Rmks
topcoders_selim	2D	tf1_keras	--	--	--	--	Empty outputs
maskrcnn	2D	tf1_keras	0.380	0.504	--	--	the model was pre-trained on DSB2018 with custom parameters
deepcell	3D	tf2	0.327	0.446	--	--	selection of only the biggest labeled object