Fig 1.
Segmentation challenges and Nessys results.
(A) Side-by-side comparison of the DAPI and LaminB1 signals obtained by confocal microscopy of pluripotent stem cells differentiated into neural cells (left) or grown in a 3D matrix (right). ‘B’ arrows indicate boundaries between touching nuclei, ‘D’ arrows point at cell debris, the ‘M’ arrow shows a mitotic nucleus, and ‘I’ arrows indicate invaginations of LaminB1 into the nucleus. (B) Confocal micrographs showing the diversity of cell shapes and volumes found in an E8.75 mouse embryo (left) and within a ‘monolayer’ of cultured cells (right). DAPI is shown in orange and LaminB1 in cyan. For each image, a magnified region is shown as a series of planes along the z axis of the image for both the DAPI channel or the LaminB1 channel. An image constructed along the yz axes is also shown. The faint vertical bar in the xy planes indicate the location of the yz image. White arrows: loss of nuclei edge in the DAPI signal, red arrow: cell debris apparent in the DAPI signal, white asterisks: flat and large nuclei distinct from their surrounding cells. (C) Images of segmented nuclei obtained with Nessys. Nuclei are assigned a unique label and a random colour. The same regions as in (B) are shown. Notice how overlapping nuclei with distinct morphology are identified accurately. E, embryonic day; Nessys, Nuclear Envelope Segmentation System.
Fig 2.
Overview of the Nessys segmentation method.
The main steps of the method are shown in white boxes annotated with a blue asterisk whenever the step is further detailed in S1 Text. Where possible, a snapshot of the intermediate output is shown. Iterations are represented with orange boxes annotated with an icon to indicate whether iterations are parallelised or sequential. Instructions that are part of the same iteration are contained within the same rounded box. In the ‘tree-structured ridge-tracing’ step, a full tree is overlaid on the image and corresponds to the diagram on the left of the image. The red circle represents the root (maximum where the procedure was initialised), smaller circles indicate the leaves of the tree, and lines represent the branches of the tree. The use of a (reusable) trained naive Bayes classifier is shown with a red box. This classifier is trained by the user before running the method. DoG, Difference of Gaussian; Nessys, Nuclear Envelope Segmentation System.
Fig 3.
Overview of the manually segmented image dataset and sample error maps created with Nessys validation and benchmarking tools.
Representative test images of the DISCEPTS image set are shown in the left column of the table (scale bars: 50 μm). For each image, red outlines represent the manually segmented image regions used to evaluate accurate hits and errors, which are shown as 3D maps in the other columns. If multiple regions are drawn, a number in a grey box indicates the correspondence between a given region and the matching 3D map. Second column: Accurate nuclei identified with Nessys. Third column: Maps of Nessys errors. Fourth and fifth column: Map of errors for the best and second-best NN. methods. The legend for these errors is indicated in the top right corner of the figure. Merge: Only one nucleus found in AM when several nuclei are present in GT. Miss: Nucleus found in GT but absent from AM. Split: Several nuclei found in AM when only one is present in GT. Spur: Nucleus generated by AM that does not exist in GT. Please note that the ‘Monolayer’ and the ‘E8.75’ images are the same images as in Fig 1B. AM, automated method; E, embryonic day; GT, ground truth; Nessys, Nuclear Envelope Segmentation System; NN., non-Nessys; Spur, Spurious.
Fig 4.
Error counting, morphological accuracy, and computational time benchmarking.
(A) Radar charts representing precision and recall for each tested method over the DISCEPTS dataset. (B) Stacked bar chart showing the proportion of detection accuracy and errors for each tested segmentation method (‘M’: Monolayer, ‘A’: Acini, ‘B’: Blastocysts, ‘7’: E7.5, ‘8’: E8.75). (C) Histograms of the deviation of morphological features from the ground truth (to create a single graph for each feature, 200 cells from each biological dataset were randomly sampled and pooled together so each dataset would contribute equally to the graph). A table summarises the mean and SD of the distribution for each tested method (‘N’: Nessys, ‘I’: Ilastik, ‘M’: MINS, ‘F’: Farsight). The vertical dashed line shows the value of the measure for the ground truth (see also S1 Text section B3). (D) Maximum image size successfully processed by the tested methods in GB. Dashes indicate that the largest tested image was successfully processed. (E) Processing times in seconds recorded for each tested method on <200-Mb benchmarking images. (F) Specifications of the computer used for benchmarking. (G) Scatter plot of the number of processed cells per seconds by Nessys versus the size of input image (E8.75 zone 1 resized by varying either plane sizes or number of planes while keeping resolution identical). The same input variables were used throughout. Data tables listing individual measurements used for the figure are available on GitLab (https://framagit.org/pickcellslab/data/2019_nessys) except for panel (G), which are given in S5 Table. E, embryonic day; GB, gigabytes; Mpx, megapixels; Nessys, Nuclear Envelope Segmentation System; RAM, random-access memory.
Fig 5.
Quantitative analysis of Tcf15 expression in a complex E8.75 mouse embryo.
(A) Scatterplot of the median Tcf15-Venus intensity within the cell versus the cell position along the Sagittal X coordinate. Boundaries between somites are apparent as gaps between groups of Tcf15-high cells. (B) Same as (A) except that points are colour-coded according to their respective embryonic region as shown in (D). (C) 3D rendering of the all detected nuclei in the E8.75 mouse embryo image. The heatmap of the relative Tcf15-venus intensity is shown only for the tailbud and somitic regions; other regions are shown in grey. (D) 3D rendering of nuclei grouped by embryonic regions. Scale bar in (C) and (D): 100 μm. The yellow dashed outline delineates the shape of the embryo. Data tables listing individual measurements used for the figure are available on GitLab (https://framagit.org/pickcellslab/data/2019_nessys). AFU, arbitrary fluorescence units; E, embryonic day; Tcf15, transcription factor 15.
Fig 6.
A novel synthetic NE fluorescent protein for live cell tracking.
(A) Schematic representing the structure of the NE live reporter and its expected topology and localisation within the cells. (B) High-resolution confocal image showing the reporter localisation within a stable cell line constitutively expressing the reporter. EMD-TM, 44 C-ter amino acids from the human Emerin protein (UniprotKB: P50402) that contain a transmembrane domain; ER, endoplasmic reticulum; (GS)3, Gly-Ser linker; NE, nuclear envelope; NLS, nuclear localisation signal.
Fig 7.
Lineage and neighbour exchange analyses reveal that cell mixing is favoured by divisions above the epithelial plane.
(A) Magnified snapshots from a single optical z slice of S3 Movie. An asterisk indicates a cell moving underneath a bright elongated cell. The arrowhead shows the trailing side of the moving cell. Time is indicated as HH:MM. (B) Single frame from the Sox1 time-lapse experiment. Individual tracks are overlaid on top of the mKate2 (green) and Sox1 (red) signals. The dots indicate the current position of the detected cells. Location of the snapshots shown in (A) is indicated by a yellow square. (C) Schematic representing the different part of a lineage tree and the colour code used for classification. (D-E) Waffle charts representing the destiny of individual lineage branch either in terms of apoptosis, survival, or division (D) or in terms of Sox1 identity (E). Each square represents one branch in the dataset. (F-G) Waffle charts comparing the destiny of sister branches for survival (F) or for differentiation (G). (H) Beeswarm-boxplots showing the z coordinate or the anisotropy of dividing and nondividing nuclei. (I) Beeswarm-boxplots showing the log2 ratio of mother branch to daughter branches of the average velocity or neighbour exchange rate. (J) Representative track (yellow) containing a division above the epithelial plane, which leads to greater neighbour exchange rate after division (white dots represent other detected cells in the image). The corresponding colour-coded trees based on z coordinate or neighbour exchange rate are shown below images (gaps in the tree correspond to time frames within which the cell was not detected). Note the high z coordinate at division and the increase in neighbour exchange rate after division. Data tables listing individual measurements used for the figure are available on GitLab (https://framagit.org/pickcellslab/data/2019_nessys). GFP, green fluorescent protein; Sox1, SRY-box transcription factor 1.