 |
This workshop is partially supported
Summary of the results
After the first call for participation, six teams registered for participation in the challenge, but only two teams successfully submitted working systems, where both methods present competitive results. As a baseline, we also present our latest method to this challenge. This page shows the quantitative evaluation of all methods.
Evaluation method
Dataset: 810 synthetic cell images from test dataset, no ground truth was given previously.
Good Segmentation: a “good” segmentation is considered to be one with Dice > 0.7 .
Abbreviations:
FNo: object-based False negative rate
TPp: pixel-based True positive rate
FPp: pixel-based False positive rate
Quantitative Comparison
The quantitative comparison contains the evaluation of cytoplasm segmentation and nuclei detection.
Table I - Nuclei
| Precision (Object) | Recall (Object) | Precision (Pixel) | Recall (Pixel) | ZSI (Pixel) | |
|---|---|---|---|---|---|
| Dani | 0.959 | 0.895 | 0.968±0.055 | 0.871±0.069 | 0.914±0.039 |
| Masoud | 0.903 | 0.893 | 0.901±0.097 | 0.916±0.093 | 0.900±0.053 |
| Lu | 0.977 | 0.883 | 0.942±0.078 | 0.912±0.081 | 0.921±0.049 |
Table II - Cytoplasm
| Dice | FNo | TPp | FPp | |
|---|---|---|---|---|
| Ushizima | 0.872±0.082 | 0.267±0.278 | 0.841±0.130 | 0.002±0.003 |
| Nosrati | 0.871±0.075 | 0.111±0.166 | 0.875±0.086 | 0.004±0.004 |
| Lu | 0.893±0.082 | 0.316±0.295 | 0.905±0.096 | 0.004±0.005 |
Ranking
SEGMENTATION OF SUBCELLULAR COMPARTIMENTS COMBINING SUPERPIXEL REPRESENTATION WITH VORONOI DIAGRAMS, Daniela M. Ushizima (Lawrence Berkeley National Laboratory, Berkeley, CA, USA), Andrea G. C. Bianchi and Claudia M. Carneiro (Federal University of Ouro Preto, Ouro Preto, MG, Brazil) [PDF]
A VARIATIONAL APPROACH FOR OVERLAPPING CELL SEGMENTATION, Masoud S. Nosrati and Ghassan Hamarneh (Medical Image Analysis Lab., Simon Fraser University, BC, Canada) [PDF]