We sought to develop a prediction workflow that converts an input image taken under phase contrast, fluorescence, or bright field into a binary mask for cell bodies. However, when microscopy is performed under different modalities, pixel intensity variations between imaging conditions make it difficult to perform semantic segmentation with a single procedure. Thus, to minimize the impact of image intensity fluctuations that inevitably arise from varying imaging sources, the input images were transformed into intermediate image representations obtained from the shape and curvature (the Hessian or second-order differentiation) of the imaged objects. This strategy is possible because in rod-shaped bacteria, the characteristic dome-shaped curvature of the poles is remarkably conserved across division cycles (Campos et al., 2014). The curvature changes in the intensity field of an image are thus represented in a so-called Shape Index map (SI) derived from the eigenvalues of the Hessian of the image (Koenderink and van Doorn, 1992) (see Materials and methods section). Therefore, all microscopy images can be transformed into SI images with pixel intensity values varying as a function of object curvature and ranging from –1 to +1, a –1 value representing a negative dome-shape and +1 representing a positive dome-shape (Koenderink and van Doorn, 1992, Figure 1A). SI images of bacterial cells acquired under various experimental conditions were used to train a U-Net.

Figure 1 with 1 supplement see all

Download asset Open asset

U-net type architectures allow fast learning from a relatively small body of labeled data because the embedded skip-connections allow the convolutional kernels between both encoder and decoder ends to be shared (Falk et al., 2019). Nevertheless, the labeled data must be representative of the broadly varying experimental conditions to produce reliable outputs: in our case, different bacterial species recorded under varying imaging modalities. A schematic of the training strategy is shown in Figure 1 and detailed in the Materials and methods section. Specifically, the U-Net was trained to segment cells by learning shapes of individual bacteria and patterns emerging from the tight contact between cells. Accordingly, we curated a hand-segmented dataset of 350 bright-field images (n = 34807 cells) of two rod-shaped bacterial species, Escherichia coli and Myxococcus xanthus. This annotated data was however insufficient for the robust segmentation of bacterial cells, thus it was further enriched with synthetic representations of rod-shaped bacteria of varying length but fixed 10 pixels width corresponding to ≈ 0.6 µm. Overall, the ground truth data used for training had two classes: one corresponding to bacteria cell bodies and the other corresponding to the contour of the detected cell (Figure 1B). This makes post-processing possible when there is not enough edge information for the proper separation of tightly connected bacterial cells, using algorithms like watershed, conditional random fields, or snake segmentation (Yang and Cao, 2013, see below).

Prior to segmentation, two parameters must be adjusted to generate a SI image from an input image. Because the width of bacterial cells was set to 10 pixels in the training data set, the input image must be scaled similarly so that the width of bacterial cells is also contained in 10 pixels. However, this scaling often smoothens the original image, which in turns smoothens the corresponding SI Image (Figure 1—figure supplement 1). This is potentially problematic because the U-Net distinguishes smooth curvatures with well-defined boundaries and noise reduction can lead to increased false positive segmentation in the scaled images. Such false positive detection is more frequent in images where the number of bacteria is sparse. This problem can be solved by adding synthetic noise to the scaled images. In total, the MiSiC workflow takes raw input images of any imaging modality, scales them and adds noise to generate SI that are then segmented with the above described U-Net (Figure 1B). For the users, scaling and potential addition of noise can be easily adjusted in the Napari GUI, which is explained in detail in a dedicated handbook.

We first tested whether MiSiC can efficiently segment images of bacteria of distinct shapes (E. coli, smaller and thicker cells and M. xanthus, longer and thinner cells) in dense colonies, captured by phase contrast, fluorescence or brightfield. To quantify the accuracy of segmentation and compare the relative performance of MiSiC for each modality, we compared the obtained MiSiC masks with hand-annotated masks (considered ground truth masks) of the same images and measured the Jaccard index (JI) as a function of the Intersection-over-Union (IoU) threshold for each modality (see Materials and methods, Figure 2a and b, Figure 1—figure supplement 1 a,b). In all cases, high JI values ( ≥ 0.8) were observed for IoU thresholds 0–0.6, indicating that MiSiC can robustly segment all modalities. In fact, MiSiC performance was comparable to the observer’s eye for each modality (or even slightly better) because similar JI scores were obtained when the same ground-truth data (generated by the same observer consistently throughout the study) was compared to data annotated by another independent observer (Figure 1—figure supplement 1c). As would be expected, the quality of the segmentation was nevertheless variable (albeit slightly) as a function of the imaging modality, the best results being obtained for fluorescence (if the fluorescence is homogenous between cells as seen in E. coli, less so for Myxococcus cells where the fluorescence levels were more variable), followed by phase contrast and bright field.

Figure 2 with 1 supplement see all

Download asset Open asset

Figure 2—source data 1

E. coli IoU over threshold -Brightfield.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig2-data1-v2.csv

Download elife-65151-fig2-data1-v2.csv

Figure 2—source data 2

E. coli IoU over threshold -Fluorescence.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig2-data2-v2.csv

Download elife-65151-fig2-data2-v2.csv

Figure 2—source data 3

E. coli IoU over threshold – Phase contrast.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig2-data3-v2.csv

Download elife-65151-fig2-data3-v2.csv

We next tested if MiSiC can segment bacterial species other than those that it was trained on (E. coli and M. xanthus), other rod shapes (Pseudomonas aeruginosa), curved shapes (Caulobacter crescentus) and filamentous shapes (Bacillus subtilis) (Figure 3a and b). For each analysis, we derived JI scores based on the comparison between MiSiC masks and hand-annotated data as described above. MiSiC predicted accurate masks for each bacterial species with JI scores ≥ 0.8 up to 0.5 IoU thresholds for all cases. Nevertheless, and as expected the segmentation accuracy was lower for curved bacteria than rod shaped bacteria (compare P. aeruginosa and C. crescentus). For filamentous bacteria, the filaments were remarkably well resolved but that was not always the case for cell separations (septa) within the filaments, likely because of insufficient edge information at cell septa in the raw image. Note that this problem may be resolved by post-processing, for example by applying a watershed algorithm to the MiSiC mask (Figure 2—figure supplement 1), which effectively resolves additional undetected septa. Although MiSiC can thus detect a number of bacterial species, a current limitation lies in its ability to segment round or oval cells. In fact, such shapes were initially omitted from the original training dataset to omit the spurious detection of round non-cell objects that are frequently observed in background images.

Figure 3 with 1 supplement see all

Download asset Open asset

Figure 3—source data 1

IoU over threshold -Bacillus subtilis.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig3-data1-v2.csv

Download elife-65151-fig3-data1-v2.csv

Figure 3—source data 2

IoU over threshold -Caulobacter crescentus.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig3-data2-v2.csv

Download elife-65151-fig3-data2-v2.csv

Figure 3—source data 3

IoU over threshold – Pseudomonas aeruginosa.

Thresholds are the same in Figure 2.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig3-data3-v2.csv

Download elife-65151-fig3-data3-v2.csv

Encouraged by these results, we tested whether MiSiC could be further used to study bacterial multicellular organization and inter-species interactions accurately at very large scales. As a model system we used Myxococcus xanthus, a delta proteobacterium living in soil, that predates collectively in a process whereby thousands of cells move together to invade and kill prey colonies (Figure 4, Pérez et al., 2016). In the laboratory, spotting a Myxococcus colony next to a prey colony (here E. coli) results in invasion and complete digestion of prey cells in 48 H (Figure 4b). To capture predator-prey interactions at single cell resolution, we set up a predation assay where Myxococcus and E. coli interact on a 1 cm² agar surface directly on a microscope slide (Figure 4—figure supplement 1a). Under these conditions, the entire invasion process occurs over a single prey cell layer allowing identification of single predator and prey cells at any given stage. This area is nevertheless quite large, and to record it with cell-level resolution, we implemented a multi-modal imaging technique termed ’Bacto-Hubble’ (in reference to the Hubble telescope and its use for the reconstruction of large scale images of the galaxies) that scans the entire bacterial community with a 100 X microscope objective and reconstructs a single image by near neighbor end-joining of multiple tiles of 80 nm/pixel resolution images (Figure 4—figure supplement 1b). Application of this method requires addressing practical considerations that are detailed in the methods section. Bacto-Hubble images (phase contrast and multi-channel fluorescence) thus capture cellular processes in a native community environment. We next tested whether MiSiC addresses the computational challenges posed by the analysis of such complex (dense population and mixed species) and large size data sets.

Figure 4 with 1 supplement see all

Download asset Open asset

Figure 4—source data 1

Fluorescence maxima – Figure 4e.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig4-data1-v2.csv

Download elife-65151-fig4-data1-v2.csv

Figure 4—source data 2

Cell length – Figure 4e.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig4-data2-v2.csv

Download elife-65151-fig4-data2-v2.csv

Figure 4—source data 3

Z-ring positioning for Figure 4h.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig4-data3-v2.csv

Download elife-65151-fig4-data3-v2.csv

First, we tested the capacity of MiSiC to segment closely-packed swarms of Myxococcus xanthus cells captured in a single image tile. To test the fidelity of segmentations in these conditions, we imaged a swarm composed of cells expressing SgmX-sfGFP, a motility protein that localizes to the cell pole (Mercier et al., 2020; Potapova et al., 2020) in both fluorescence and phase contrast modalities (Figure 4c). Phase contrast images were used to obtain a MiSiC segmentation mask (Figure 4d). Subsequently, the mask was processed under MicrobeJ (Ducret et al., 2016) to remove objects that do not correspond to cells (Materials and methods, less than 1.4%, n = 1695) and calculate the localization pattern of SgmX-GFP foci with respect to the long axis of each segmented cell (Figure 3e). As expected, SgmX-GFP loci localized to a cell pole, consistent with most Myxococcus cells in swarms being properly segmented by MiSiC.

Second, to show that MiSiC can be used to quantitatively study cellular processes in entire Bacto-Hubble images, we mapped a Myxococcus cellular process directly during prey invasion. Cell division is expected to occur mostly in prey-areas in absence of any other source of nutrients. Like all rod-shaped bacteria, dividing Myxococcus cells assemble a polymeric FtsZ bacterial tubulin ring to initiate cell division (Schumacher et al., 2017). When it is fused to fluorescent proteins, the FtsZ ring is observed as a dot at mid-cell (Treuner-Lange et al., 2013), which can be used as a proxy to determine which cells enter division. Thus, we first engineered M. xanthus cells expressing FtsZ fused to Neon-Green (NG, Materials and methods) and mixed Myxococcus FtsZ-NG+ (5%) with non-labeled cells (95%) in the presence of an Escherichia coli prey cell colony. A fluorescence Bacto-Hubble image spanning ~5 mm² (representing 225 tiles of 500 × 500 pixels images) of the community during the invasion phase (Figure 4f) was then captured and segmented tile-by-tile using MiSiC (Figure 4f–g). Due to the fact that MiSiC uses SI, it is remarkably insensitive to noise and intensity variations between images (Figure 4—figure supplement 1c,d) and thus it easily allows segmentation of multi-tile images where signal intensity variations between tiles is to a large extent unavoidable. Cells with mid-cell FtsZ-NG fluorescence clusters were clearly observed suggesting that cell division is ongoing (Figure 4h). Dividing cells were therefore counted across the entire image (Figure 4i, Materials and methods) to determine where they localize spatially within the community. Figure 4j, shows that cell division is markedly increased in the prey area, demonstrating directly that Myxococcus grows during prey invasion. Thus, MiSiC is appropriate for the automated detection of cellular processes (detected at subcellular resolution) at community scales.

Third, we explored how MiSiC could be further adapted to segment and classify multiple bacterial species intermingling and interacting in space; here Myxococcus cell groups invading the tightly-knitted E. coli prey colony. To segment each bacterial species directly from unlabeled phase contrast-images, new training datasets were produced and used to retrain the U-NET. These labeled datasets were obtained by imaging GFP-labeled Myxococcus and mCherry-labeled E. coli (see Materials and methods). Images were captured for each channel (GFP, mCherry) and segmented separately with MiSiC to obtain masks for each species (Figure 5a). These masks were used to retrain the U-NET, the predictive value of which was tested by deriving classification masks directly from a phase-contrast image of a mixed Myxococcus (GFP) - E. coli (mCherry) population (Figure 5). Comparison of the MiSiC classification to hand annotated GFP (Myxococcus) and mCherry (E. coli) images over selected fields where each species interact (n = 4, ie Figure 5b) respectively yielded JI scores of 0.95 ± 0.036 (n = 200 cells) and 0.89 ± 0.047 (n = 545 cells), suggesting that the classification is highly accurate despite the tight interactions between Myxococcus and E. coli cells. To further show that species classification is reliable in a large dataset, we next classified Myxococcus from E. coli cells directly in a phase contrast Bacto-Hubble image (840 tiles, 2.1.10⁹ pixels). We could thus discriminate cells belonging to each species in the predation area (Figure 5c) for a total detection of ~402,000 Myxococcus and ~630,000 E. coli cells in the entire image. Given the large size of the dataset, it would be impossible to test the accuracy of the classification procedure exhaustively by comparing it to the ground truth, instead, we tested whether the distribution of shape descriptors, such as the extent (E = area/bounding box area), solidity (S = area/convex area) and minor axis length, matched the distribution of these descriptors obtained from images of each single bacterial species (Figure 5d). The observed distributions were indeed consistent with an efficient separation of species in the mixed community, suggesting that the classification is also robust in large images. Inevitably, infrequent ambiguous classifications arise in the areas where the cells interact tightly due to low contrast in these areas. These instances generally appear as bi-color objects because the prediction is not homogeneous inside such objects. Given that Myxococcus cells and E. coli cells have clear morphometric differences, these uncertain cells can be easily eliminated by filtering with discriminating parameters (i.e. each of them or combined, Figure 5d).

Figure 5

Download asset Open asset

Figure 5—source data 1

E. coli only for Figure 5d.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig5-data1-v2.csv

Download elife-65151-fig5-data1-v2.csv

Figure 5—source data 2

E. coli in mixed colony for Figure 5d.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig5-data2-v2.csv

Download elife-65151-fig5-data2-v2.csv

Figure 5—source data 3

M.xanthus only for Figure 5d.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig5-data3-v2.csv

Download elife-65151-fig5-data3-v2.csv

Figure 5—source data 4

M.xanthus in mixed colony for Figure 5d.

https://cdn.elifesciences.org/articles/65151/elife-65151-fig5-data4-v2.csv

Download elife-65151-fig5-data4-v2.csv

The complete list of the strains used for the study is compiled in Table S1 Supplementary file 1. For predation assays, cells of Myxococcus xanthus (DZ2, Table S1, Supplementary file 1) were grown overnight in 100 mL flasks in 10–20 mL of CYE (Bustamante et al., 2004) media without antibiotics at 32°C with shaking. In parallel a colony of Escherichia coli (MG1655, Table S1) was grown in 5 mL LB medium in a glass tube at 37°C with shaking. The next day, OD_{600 nm} were measured and cultures of both strains were washed twice at 5000 rpm in CF (Bustamante et al., 2004) minimal media to discard CYE and LB traces. After the washes, the density of the cultures was brought to 5 OD units in CF media. Pads of CF agar 0.5% were poured in precast frames (in situ GeneFrame, 65 µL, ABGene, AB-0577) that were mounted on glass slides and briefly dried. One µL of both Myxococcus xanthus and prey cell suspensions were spotted as close as possible to one another on the pad making sure that they would not merge. Glass slides were kept in a sealed humid dish for 6, 24, 48, or 72 hr at 32°C. Thirty min before observation, the agar pad around the colonies was cut out and discarded and the pad was sealed with a cover slip and observed by microscopy.

Bacillus subtilis strains were grown in LB medium at 37°C until they reached an OD_{600 nm} of 0.6, were transferred into wells and covered in a low melting LB-based agarose suspension (2%) before observation. Pseudomonas aeruginosa was grown in LB medium and cells were observed at an OD_{600 nm} of 0.5. Caulobacter crescentus was cultured in PYE on the benchtop without shaking for 3–4 days. Cells were imaged from disrupted fragments of a pellicle biofilm and were transferred from the air-liquid interface to glass slides for imaging (Marks et al., 2010). Anabaena nostoc was grown in BG11 medium at 30°C with illumination (40 µE m-2s-1). Finally, Desulfovibrio vulgaris cells were grown until mid-exponential phase (OD_{600 nm} of approximately 0.4–0.5) in LS4D medium supplemented with 1 g/L of yeast extract (LS4D-YE) at 33 °C in an anaerobic chamber (COY Laboratory Products) filled with a 10% H₂-90% N₂ mixed-gas atmosphere. Cultures (200 µL) were centrifuged, and the pellet was resuspended in 100 µl of 10 mM Tris-HCl (pH 7.6), 8 mM MgSO₄ and 1 mM KH₂PO₄ buffer (TPM buffer). The cells were placed between a coverslip and an agar pad containing 2% of agarose.

To follow cell cycle progression in single cells of Myxococcus xanthus, a merodiploid strain of DZ2 expressing both native FtsZ and the fusion protein FtsZ-neonGreen (FtsZ-NG, Table S3 Supplementary file 3) was built (DM14, Table S1, Supplementary file 1). To do so, the coding sequence of DZ2 ftsZ gene (MXAN_5597) along with its predicted promoter sequence was amplified by PCR with primers oDM1 and oDM2 (Table S2, Supplementary file 2) and cloned in the non-replicative plasmid pKA32 (Treuner-Lange et al., 2013) allowing for its site-specific integration at the DZ2 attmx8 phage attachment site on the M. xanthus chromosome. The coding sequence of the neonGreen protein was amplified from a plasmid (Shaner et al., 2013) using primers oDM16 and oDM17 (Table S2 Supplementary file 2) allowing the in frame addition of neonGreen at the C-terminus of the FtsZ protein. When grown in CYE rich medium, DM14 did not present any significant defect in growth rate or cell shape. DM14 cell size is not significantly different from that of the isogenic wild type DZ2 strain. DM14 cells were spotted on thin CF agar pads to follow FtsZ localization in axenic cultures and allowed us to confirm that cell cycle progression was accompanied with the relocalization of Ftsz-nG from being diffuse in the cytoplasm to forming a discrete fluorescent focus at mid-cell before cell septation as previously described (Treuner-Lange et al., 2013).

To generate Dataset 2 (see below), strains of E. coli (EC500, Table S1 Supplementary file 1) and M. xanthus (DM31, Table S1 Supplementary file 1) expressing soluble versions of mCherry and sfGFP fluorophores respectively were used. To generate DM31, a plasmid allowing for the constitutive expression of sfGFP was built (pDM14 Table S3 Supplementary file 3). To obtain a high and constitutive expression of sfGFP in M. xanthus, we sought for the closest homolog of the constitutively expressed E. coli EF-TU (Translation elongation factor) in M. xanthus genome which is MXAN_3068. The 1000 bp region upstream of MXAN_3068 (p3068) was amplified by PCR using oDM53 and oDM54 and cloned upstream the coding sequence of sfGFP (amplified using primers oDM61 and oDM62) in a pSWU19 plasmid. The transcriptional fusion was then integrated on DZ2 chromosome at the attmx8 site through transformation. DM31 cells display a constitutive bright diffuse fluorescent signal in our growth conditions.

All microscopy images were acquired with an inverted optical microscope (Nikon TiE) and a 100 x NA = 1.45 Phase Contrast objective. Camera used was Orca-EM CCD 1000 × 1000 (Hamamatsu) camera mostly set with binning 2 × 2. Acquisition software was Nikon NIS-Elements with specific module JOBS. Fluorescence acquisition used a diode-based excitation device (Spectra-X Lumencore).

The Bacto-Hubble image is a composite image of rasters of the entire area and requires that the scanning speed must be sufficiently fast to avoid image shifts due to ongoing cell dynamics. To minimize the shift in focus from tile to tile we used Nikon perfect focus system (PFS) equipped with servo-control of the focus with an infrared LED. This is especially challenging because continuous focus alignment of the microscope slows down the acquisition times dramatically. To obtain a satisfactory compromise allowing both fast scanning and correct focusing we: (i) reduced the number of dynamic elements on the microscope set up: we replaced shutters by Light Emitting Diodes (LEDs, Spectra-X for fluorescence source and a white diode for transmitted light) which could be switched with a high-frequency rate (100 kHz). In addition, a double band dichroic mirror for the fluorescent cube was used to avoid switching the filters’ turret for each snapshot. (ii) used an EM-CCD camera set to a 2 × 2 binning mode to reduce the size of images (500 × 500 pixels at 0.16 µm/px) and acquisition time, and (iii) sped up the vertical movements by means of a piezoelectric stage. In its largest scanning mode, Bacto-Hubble thus captures 80 × 40 raster images covering a total surface of 20 mm2 (containing up to 0.8 billion pixels, an acquisition-time up to 4 hr), enabling a continuous magnification display from eye visible structures to single-cells. Individual tiles for Bacto-Hubble images were acquired with the scan large field capabilities of NIS-Elements software. A key point for Bacto-Hubble large images was the quality of the sample slide mounting. The samples were placed on a glass slide with a thin double-sided sticky frame (in situ GeneFrame, 65 µL, ABGene, AB-0577). An agar pad was poured inside the frame and a microliter of cells was placed on it. The chamber was closed with a glass cover slide. This assembly allows very good flatness and rigidity.

We implement a U-net inspired in encoder-decoder architecture with skip-connections and use it as a base network for segmentation tasks. This architecture is now widely used for segmentation tasks and has many advantages that are discussed in previous articles (Falk et al., 2019; Ronneberger et al., 2015). The original U-net architecture (Falk et al., 2019) was modified to include relu activation for all layers except for the output layer where sigmoid activation was used. The general U-net was implemented in Python programming language using tensorflow (https://www.tensorflow.org/tutorials/images/segmentation), where the number of input channels (say, n) and output classes (say, m ) could be varied as required by different models. The number of encoder layers were fixed to four with filter lengths [32,64,128,256] for the encoder side. The loss function was also modified from the original implementation. A combination binary cross entropy and the Jaccard, 1912 index was used as the loss function with Adams optimizer (learning rate = 0.001) for the minimization.

For brevity, we denote the network as a mapping between input X and output y as,, where X is a set of images with n channels and y is the output with m classes. Thus given a training data set of X_train with sizes (N × S × S× n) and multiclass images y_train (of size N × S × S× m), the generalized implementation of U-NET learns to predict the segmented image from unknown images X.

Two training datasets were used for this study. Dataset one to segment bacterial cells (MiSiC) and Dataset two to classify Myxococcus and E. coli cells in mixed cultures.

$\sigma$The shape index (SI) map of an image, x, calculated over a scale , is defined as

where k₁, k₂ (with k₁> k₂) are the eigenvalues of the Hessian of the image,$x$, calculated over a scale $\sigma$(Koenderink and van Doorn, 1992). The SI map remains within the range [−1, 1] and preserves the MiSiC shape information while being independent of the intensity values of the original image. Using the Dataset 1, we pre-process the input images X′ to generate a train set: X_train of size 1260 × 256 × 256 × 3. Each channel in X_train is the shape-index map calculated at three different scales [1,1.5,2] to obtain shape index information of the cells at various scales. An instance of the U-net is trained over this data set to produce a network able to map data represented by X_train → y_train. The network learns to reject the noise in the shape-index map and produces masks and boundaries of the cell like structures in the shape-index map. The trained network was tested over 87 cases of labelled bright-field images, that were previously unseen by the network leading to a segmentation accuracy of 0.76 computed with Jaccard, 1912 coefficient.

Preprocessing the input image to enhance the edge contrast and homogenising intensities helps in obtaining a good segmentation via MiSiC. Some of the preprocessing that gave good results are gamma correction for homogenising, unsharp masking for sharpening the image and sometimes a gaussian of laplace of the image that removes intensity variations in the entire image and keeps edge-like features.

The dataset one used to train the MiSiC contains cells with a width in the range of 8–10 pixels.Thus, to obtain a satisfactory segmentation, the input image must be scaled so that the average bacteria width is around 10 pixels. However, the scaling often modifies the original image leading to a smoother shape index map. Since, MiSiC has basically learnt to distinguish between smooth curvatures with well-defined boundaries from noisy background. A smooth image without inherent noise leads to a lot of false positive segmentations. Therefore, counterintuitively, synthetic noise must be added to the scaled or original image for a proper segmentation. It must be kept in mind that the noise variance should not reduce the contrast of the edges in the original image while it should be enough to discard spurious detections. Gaussian noise of a constant variance may be added to the entire image or alternatively, the variance could be a function of the edges in the input image.

A U-NET was trained on dataset two to segment a single channel phase-contrast image into an image containing semantic classification of each species. The probability map for each label is color coded (blue = E. coli. coli, orange = M. xanthus) such that each pixel has a probability value to be part of a given class. In rare instances, bi-color objects are obtained because in these cases the prediction is not homogeneous inside the predicted objects. For subsequent analysis, these objects were filtered for the morphometric analysis shown in Figure 5c. A trained model and a python script illustrating how it was used is available at: https://github.com/pswapnesh/MyxoColi.

The validation of the automated MiSiC results shown in Figures 2 and 3 and Figure 2—figure supplement 1, was performed by comparing MiSiC masks to hand-annotated masks obtained by a single observer, herein referred to as ground-truth. This ground truth was also compared to data annotated by a second observer to test variability between observers. For comparison, curves of Jaccard Index vs. IoU threshold were constructed based on Jeckel and Drescher, 2021. Such curves allow a precise estimate of the accuracy and quality of shape detection and prediction.

For this, the $IoU=\frac{X\cap Y}{X\cup Y}$ (X = Ground Truth object; Y = Predicted object ) was determined by overlaying each predicted MiSiC mask with its corresponding ground truth mask and the number of True Positives (TP), False Positives (FP) and False Negatives (FN) that remain above varying IoU thresholds, applied from values ranging from 0.00 to 1.00 in 0.01 intervals, was determined so as to calculate the Jaccard index defined by the ratio $JI=\frac{TP}{TP+FP+FN}$ for a given threshold value.

To estimate the accuracy of classification, JIs were determined directly by comparing hand-annotated (ground truth) datasets obtained from fluorescence image of GFP (Myxocococus) and mCherry (E. coli).

Analysis of MiSiC performance in the presence of noise and comparison with Supersegger.

To illustrate MiSiC performances in the presence of noise and in comparison with SuperSegger (Stylianidou et al., 2016), datasets consisting of 141 E. coli microcolony images were retrieved from the SuperSegger website. These images were analysed with the provided parameters with SuperSegger and with the following parameters with MiSiC: Cell width = 9, Scaling factor = Auto, Noise = 0.001, Unsharp = 0.6 and Gamma = 0.1. To assess the segmentation robustness to noise for each program, datasets were normalized by the maximum intensity value recorded in the first frame of the dataset and Gaussian noise was added with varying variance. The resulting datasets were then analysed with the initial parameters used to compute reference segmented images except for MiSiC where the noise parameter was set to 0. The relative performance of each program was then evaluated by computing Dice indices (Zijdenbos et al., 1994).

Classic morphological features: Area, Perimeter, Bounding Box (Width, Height), Circularity, Feret diameter, minimum Feret diameter, MajorAxis (ellipse), MinorAxis (ellipse), (n) number of objects.

Special calculated morphological features: Solidity = Area / Convex Area; AR = MajorAxis / MinorAxis; Extend = Area / (Width*Height).

To construct the plot shown in Figure 4e (demograph), the cells bodies were obtained with MiSiC segmentation and the binary mask was analyzed with the MicrobeJ software (Ducret et al., 2016), with a cell model set to parameters (area >0.5 µm², Circularity <0.8, ‘poles’ = 2) to filter all remaining segmented objects that do not correspond to cells. The localization of the centroid (cell middle) and length of longitudinal axis was then determined for each cell under MicrobeJ. The fluorescent clusters were detected with a local maxima filter and their position relative to the middle of the cell was plotted along the axis with a positive sign. Negative sign clusters are therefore the manifestation of rare cells with bi-polar foci. The fluorescent clusters are plotted as dots with a color scale based on spatial density.

In Figure 4i, clusters of fluorescent protein FtsZ-NG were used as cell division markers. The clusters were detected by local maxima detection (scikits-image.peak_local_max(image = fluorescence image, label = MiSiC mask, num_peaks = 1)).

The cell division ratio in Figure 4j was calculated using the spatial 2D density derived from (i) the mask of the total fluorescent cell population across the entire image and (ii), the mask of the total number of fluorescent maxima (reflecting dividing cells) across the entire image. Spatial densities were calculated with sklearn.neighbors.KernelDensity(), with a bandwidth of 1% of the image width. The proportion of dividing cells was obtained by dividing the spatial density maps: density of fluorescent maxima/density of total cells.

A MiSiC pip installable python package is available at https://github.com/pswapnesh/MiSiC (Swapnesh, 2021; copy archived at swh:1:rev:45f659124a8e207f78296d77664fb96de5472708).

A Graphic user Interface pip installable as python package is available at https://github.com/leec13/MiSiCgui (Espinosa, 2021; copy archived at swh:1:rev:97401bedd44aa7b28d22f8cf87e76b521f15f40a).

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We thank Anke Treuner-Lange in Lotte Sogaard-Andersen lab for sharing the plasmid pKA32 with us (Treuner-Lange et al., 2013). MiSiC robustness was tested on microscope images showing various bacterial strains that were kindly shared with us. For that, we thank Anne Galinier and Thierry Doan (Bacillus subtilis), Christophe Bordi (Pseudomonas aeruginosa), Aretha Fiebig (Caulobacter crescentus) and Romain Mercier (SgmX-GFP images of Myxococcus xanthus). We also thank Baptiste Piguet-Ruinet and Célia Jonas for constructing the pDM14 plasmid during their internship. We thank Jean Raphael Fantino for the design of the MiSiC Logo. SP, LE, and TM are funded by CNRS within a 80-Prime initiative. TM is funded by an ERC Advanced Grant (JAWS EM is funded by an AMIDEX-PhD program from Aix-Marseille University). MN, JBF, AL and SR are funded by two ANR grants IBM (ANR-14-CE09-0025-01) and HiResBacs (ANR-15-CE11-0023).

© 2021, Panigrahi et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.