Table 1. Comparison of segmentation performance among different segmentation algorithms and datasets.
The Jaccard (JI) and Dice (DI) indices were computed on a validation data set as previously described [53]. The Jaccard index for E. coli was estimated from the segmentation error rate (0.5% in this work and 2.6% in [17]) assuming each segmentation error was due to the incorrect joining of 2 bacteria. NA—Not available.
| Cell type(s) | Method | Imaging type | JI | DI | Dense | Processing time per image | Curation time per image | Citation |
|---|---|---|---|---|---|---|---|---|
| Bacteria | ||||||||
| Coliform bacteria | Edge finding and thresholding | Phase | ~0.94 | Yes | 5 s | ~10–30 s | [15,14] | |
| Coliform bacteria | Conv-nets | Phase | ~0.95 | Yes | 120 s | 1–2 s | This work | |
| Mammalian nuclei | ||||||||
| Thresholding and level sets | Fluorescence | 0.88 | Yes | 2–3 s | ~ 5–10 s | [45] | ||
| Conv-nets and thresholding | Fluorescence | 0.89 | 0.94 | Yes | 10 s | ~0–1 s | This work | |
| Mammalian cytoplasm | ||||||||
| N1E115 | Voronoi diagrams | Fluorescence | 0.75 | No | 40 s | NA | [46] | |
| Neural stem cells | Level sets | Bright field | 0.79 | No | NA | NA | [47] | |
| NIH-3T3 | Edge finding, thresholding, and morphological operations | Phase | 0.76–0.92 | No | NA | NA | [21] | |
| HeLa-S3 | Ilastik (texture/edge filters and random forest classifier) | Phase | 0.56* | 0.71* | Yes | 1s | NA | [18], This work |
| HeLa-S3 | Conv-nets | DIC | 0.77 | Yes | 1s | NA | [37] | |
| U373 | Conv-nets | DIC | 0.93 | No | 1s | NA | [37] | |
| MCF10A | Conv-nets and active contours | Phase | 0.77 | 0.86 | Yes | 75 s | ~ 2–10 s | This work |
| NIH-3T3 | Conv-nets and active contours | Phase | 0.77 | 0.86 | Yes | 75 s | ~ 2–10 s | This work |
| HeLa-S3 | Conv-nets and active contours | Phase | 0.71**-0.84 | 0.83**-0.92 | Yes | 75 s | ~ 2–10 s | This work |