Benchmark for multi-cellular segmentation of bright field microscopy images
1 Blavatnik School of Computer Science, Tel Aviv University, Tel Aviv, 69978, Israel
2 School of Physics and Astronomy, The Raymond and Beverly Sackler Faculty of Exact Sciences, Tel Aviv University, Tel-Aviv, 69978, Israel
3 Center for Theoretical Biological Physics, Rice University, Houston, TX, 77005-1827, USA
4 Research & Development Unit Assaf Harofeh Medical Center, Zerifin, 70300, Israel
5 Department of Clinical Microbiology and Immunology, Sackler School of Medicine, Tel Aviv University, Tel Aviv, 69978, Israel
BMC Bioinformatics 2013, 14:319 doi:10.1186/1471-2105-14-319Published: 7 November 2013
Additional file 1:
Supporting Text. This file contains a brief description of the evaluated algorithms, notes on parameter tuning, details on evaluation of Topman’s thresholding method, and details on assessing the baseline variance in the annotated data.
Format: PDF Size: 496KB Download file
This file can be viewed with: Adobe Acrobat Reader
Additional file 2: Table S1:
Precision/recall. Precision/recall of all algorithms on all datasets.
Format: PPTX Size: 43KB Download file
Additional file 3: Figure S1:
Direct comparison of algorithms on all images. Image-by-image evaluation. Scatter plots displaying for each image the F-measure produced by the 3 algorithms. Each x-axis entry represents an image (ordered by the filename), y-axis is the F-measure. Red – Tscratch, Green – MultiCellSeg, Cyan – Topman’s algorithm. a, Init. b, NN15. c, Melanoma. d, TScratch. e, Scatter. f, Microfluidics. g, HEK293. h, MDCK.
Format: TIFF Size: 2.3MB Download file
Additional file 4: Table S2:
Baseline variance. An arbitrary partial set of the images (62 images from all datasets, excluding the “Scatter” dataset) was selected to be annotated by another expert. This annotation was compared with the primary annotated ground truth by calculating the mean F-measure to assess the baseline variance of each dataset.
Format: PPTX Size: 41KB Download file
Additional file 5: Figure S2:
Baseline variance examples. Visualization of inconsistencies between manual annotations by different experts. Annotations shown were selected from the dataset with higher baseline variance (“Melanoma”, “Miscrofluidics”). The green channel is the raw image, the blue channel is the official annotation of cells, and the red channel is the second annotation. Thus, light-magenta represents agreement in annotation of cells, green represents agreement in annotation of non-cellular regions, light-red represents regions annotated as non-cellular in the ground truth but as cellular by the second expert, light blue represents regions that were annotated as cellular according to the ground truth but non-cellular according to the second expert. It is clear from this visualization that most inconsistencies appear at cell borders.
Format: TIFF Size: 5.1MB Download file
Additional file 6: Table S3:
Adjusting Tompan’s algorithm. The automatic threshold extraction method in Topman’s algorithm was evaluated compared to a constant threshold. Evaluation of different values demonstrated that a constant threshold surpasses the automatic adjustment for most datasets. The best value found was used to evaluate this algorithm’s performance in the main text.
Format: TIFF Size: 734KB Download file