Fig 1.
Typical barriers when using segmentation software.
Cell segmentation methods are typically designed for specific applications and therefore often lack a data management system with a versatile data importer. This shortcoming results in the need for file format and image shape conversion steps. Furthermore, multiple tools often need to be combined to cover the whole workflow, from training data creation to applying trained models. Again, further processing steps may be required to enable tool compatibility. For many applications, it is not (yet) possible to do without own annotated training data. Interactive cropping functionalities are helpful in this case and enable an efficient annotation for dense-growing organisms. Nevertheless, this feature has not yet been included in cell segmentation software. Note: fluorescence images have been inverted for Omnipose [13], and phase contrast images have been inverted for Cellpose [14] for better results.
Fig 2.
OMERO is used for data management since it provides a versatile data importer and standardizes file handling [15]. Data can be viewed in the browser with the OMERO.web client. microbeSEG offers training data creation, training, evaluation, and inference functionalities. The jointly developed toolkit ObiWan-Microbi is used for manual annotation and result correction [18].
Fig 3.
Automatically proposed crops can be selected and uploaded to OMERO (a). The crop proposals are extracted randomly from different non-overlapping image regions (the left crop originates from the left image region, and the right crop from the right). For the pre-labeling, it is possible to upload only the image or the image with its prediction (b). The image-only upload is helpful when the pre-label predictions require too many manual corrections.
Fig 4.
Training data representations of the two microbeSEG methods.
From the ground truth (a—instances: color-coded), the boundary representation (b—cell interior: gray, cell boundary: white) and the distance representations (c—distance to background, inverse distance to neighbors) can be computed. A deep learning model is trained to predict either the boundary representation or the distance representations, and the single instances are recovered in the post-processing.
Table 1.
microbeSEG accuracy in dependence of annotation time.
All methods are evaluated on 12 B. subtilis and 12 E. coli test images. For each microbeSEG setting, the median aggregated Jaccard index AJI+ out of five trained models is shown. The times include the crop selection.
Fig 5.
Exemplary images and segmentation overlays for B. subtilis and E.coli.
Shown are the results of the median distance method microbeSEG models from Table 1. S2 Fig shows the raw network outputs of the microbeSEG model (d) and results for the boundary method.
Fig 6.
Segmentation of U2OS cells from the BBBC039 dataset [32].
The microbeSEG segmentation model (default settings) has been trained on HeLa cells from the Cell Tracking Challenge [33, 34].
Fig 7.
Exemplary microbeSEG segmentation (a) and analysis results (b) for a growing C. glutamicum colony.
The results can easily be viewed in Fiji. S1 Video shows a video of the segmentation results of the growing colony.
Table 2.
Cell segmentation software key feature comparison.
Considered are only tools with a graphical user interface since end users should not need programming expertise. Non deep learning segmentation methods may require expert knowledge for parametrization and are not state-of-the-art anymore. Data format support does not necessarily mean that each image can be processed: if no data management system (DMS) with metadata support is used, e.g., the channel dimension can be the first or the last dimension for.tif files, and the method may have requirements on the channel dimension position. ━: feature not fulfilled/supported, : feature only fulfilled/supported with restrictions, ✔: feature fulfilled/supported.