Active mesh and neural network pipeline for cell aggregate segmentation

Abstract

Segmenting cells within cellular aggregates in 3D is a growing challenge in cell biology due to improvements in capacity and accuracy of microscopy techniques. Here, we describe a pipeline to segment images of cell aggregates in 3D. The pipeline combines neural network segmentations with active meshes. We apply our segmentation method to cultured mouse mammary gland organoids imaged over 24 h with oblique plane microscopy, a high-throughput light-sheet fluorescence microscopy technique. We show that our method can also be applied to images of mouse embryonic stem cells imaged with a spinning disc microscope. We segment individual cells based on nuclei and cell membrane fluorescent markers, and track cells over time. We describe metrics to quantify the quality of the automated segmentation. Our segmentation pipeline involves a Fiji plugin that implements active mesh deformation and allows a user to create training data, automatically obtain segmentation meshes from original image data or neural network prediction, and manually curate segmentation data to identify and correct mistakes. Our active meshes-based approach facilitates segmentation postprocessing, correction, and integration with neural network prediction.

Figure 1

Overview of segmentation pipeline, from an original two-channel 3D fluorescent microscopy image to a set of meshes that represent the cell nuclei and the cell membranes. To see this figure in color, go online. For a Figure360 author presentation of this figure, see https://doi.org/10.1016/j.bpj.2023.03.038.

Figure 2

x-y cross sections through the equator of six different organoids after 8 h of imaging. Scale bar, 10 $\mu \mathrm{m}$. Red label, membrane dye; magenta, DNA. Organoids in (A–F) are later referred to as Movies 1–6, corresponding to Video S1. Segmentation result for cell membrane and cell nuclei for Movie 1, view removing centroid motion, Video S2. Segmentation result for cell membrane and cell nuclei for Movie 2, view removing centroid motion, Video S3. Segmentation result for cell membrane and cell nuclei for Movie 3, view removing centroid motion, Video S4. Segmentation result for cell membrane and cell nuclei for Movie 4, view removing centroid motion, Video S5. Segmentation result for cell membrane and cell nuclei for Movie 5, view removing centroid motion, Video S6. Segmentation result for cell membrane and cell nuclei for Movie 6, view removing centroid motion. To see this figure in color, go online.

Figure 3

Manual initialization of segmentation meshes that are then deformed using the active mesh method to the cell nucleus (A–C) or to the cell membrane (D–F). (A and D) Orthogonal cross section views and a 3D view during mesh initialization. Red circles: boundaries of the spheres used for mesh initialization. The yellow and blue circles are handles that can be manipulated by the user to adjust the position and radius of the spheres. (B and E) Same orthogonal views with the initialized mesh. (F and G) Mesh after deformation to the nucleus or cell membrane image intensity. To see this figure in color, go online.

Figure 4

Analysis of automated segmentation quality. (A and B) Scatterplot of best Jaccard index (JI) versus the distance between the ground truth center of mass and the predicted center of mass ($\Delta$CM) for cells from a “seen” and an “unseen” data set. The filled circles represent the mean values of the data points, and the error bars reflect the standard deviation. (A) Results of automated segmentation of cell nuclei at full resolution (voxels with side length 0.175 μm). A nucleus diameter is about 8 μm. (B) Results of automated segmentation of cell membrane at full resolution. Insets: histogram of best JI distributions. Individual data points outside of the plot range: (A) 1/300, (B) 2/300. To see this figure in color, go online.

Figure 5

Segmentations results for cell membrane and cell nuclei for six different mammary gland organoids, segmented over 24 h of growth. (A–F) Within each box, segmentation meshes are shown at 0, 8, 16, and 24 h for each organoid. Within each box, top row: solid volumes correspond to nuclei segmentation meshes and wireframes to cell membrane segmentation meshes. Bottom row: example trajectory of a cell nucleus and the nuclei of the cell progeny during the video. To see this figure in color, go online. See Video S1. Segmentation result for cell membrane and cell nuclei for Movie 1, view removing centroid motion, Video S2. Segmentation result for cell membrane and cell nuclei for Movie 2, view removing centroid motion, Video S3. Segmentation result for cell membrane and cell nuclei for Movie 3, view removing centroid motion, Video S4. Segmentation result for cell membrane and cell nuclei for Movie 4, view removing centroid motion, Video S5. Segmentation result for cell membrane and cell nuclei for Movie 5, view removing centroid motion, Video S6. Segmentation result for cell membrane and cell nuclei for Movie 6, view removing centroid motion, Video S7. Segmentation result for cell membrane and cell nuclei for Movie 1, view removing centroid motion and following solid rotation of the organoid, Video S8. Segmentation result for cell membrane and cell nuclei for Movie 2, view removing centroid motion and following solid rotation of the organoid, Video S9. Segmentation result for cell membrane and cell nuclei for Movie 3, view removing centroid motion and following solid rotation of the organoid, Video S10. Segmentation result for cell membrane and cell nuclei for Movie 4, view removing centroid motion and following solid rotation of the organoid, Video S11. Segmentation result for cell membrane and cell nuclei for Movie 5, view removing centroid motion and following solid rotation of the organoid, Video S12. Segmentation result for cell membrane and cell nuclei for Movie 6, view removing centroid motion and following solid rotation of the organoid.

Figure 6

Cross section and 3D view for one frame of one mammary gland organoid shown in Fig. 5. The cross sections display overlay of nuclei (filled volumes) and membrane (wireframes) segmentation meshes on the original data (red, membrane dye; gray, DNA label). To see this figure in color, go online.

Figure 7

Quantifications associated with tracked nuclei for the six segmented organoids. (A) Probability distribution of nucleus velocity, for each individual video. (B) Number of cells as a function of time. To see this figure in color, go online.

Figure 8

Analysis of automated segmentation quality at half resolution, with a larger training data set. (A and B) Scatterplot of best JI versus the distance between the ground truth center of mass and the predicted center of mass ($\Delta$CM) for cells from the same “seen” and “unseen” data sets as in Fig. 4 (here the training data set is larger than the seen data set). The filled circles represent the mean of the data points, and the errorbars reflect the standard deviation. (A) Results of automated segmentation of cell nuclei at half resolution (0.350 μm voxels). (B) Results of automated segmentation of cell membrane at half resolution. Insets: histogram of best JI distributions. Individual data points outside of the plot range: (A) 0/300, (B) 10/300. To see this figure in color, go online.

Figure 9

Analysis of segmentation quality with StarDist. (A and B) Scatterplot of best JI versus the distance between the ground truth center of mass and the predicted center of mass ($\Delta$CM) for cells from a “seen” and an “unseen” data set. The solid circles represent the mean value of the data points and the errorbars reflect the standard deviation. (A) Results of automated segmentation of cell nuclei. (B) Results of automated segmentation of cell membrane. Insets: histogram of best JI distributions. (C) Representative example of nucleus prediction from StarDist for two different planes of view. (D) Representative example of membrane prediction from StarDist, for two different planes of view. In (C) and (D), gray regions correspond to StarDist-predicted labels, colored lines indicate ground truth segmentation meshes. Individual data points outside of the plot range: (A) 2/300, (B) 7/300. To see this figure in color, go online.

Figure 10

Mouse embryonic stem cell colony imaged on a spinning disc confocal microscope. (A) Cross sections of original image (left) with ground truth segmentation result overlaid (right). White, nuclear label; red, membrane label; other colors, contours of membrane segmentation meshes and filled regions of nuclei segmentation meshes. (B) Cross section of neural network output, before and after training the network on the spinning disc images. Top images: nuclei segmentation; bottom images: membrane segmentation. Colors correspond to different outputs of the neural network. Green, mask; red, border; blue, distance transform. Green mask label indicates background. (C and D) Scatterplot of best JI versus the distance between the ground truth center of mass and the predicted center of mass ($\Delta$CM), before (“untrained”) and after (“trained”) training of the network on two frames of a video of the colony. The filled circles are the mean values for the respective datasets and the error bars are one standard deviation.(C) Results of automated segmentation of cell nuclei. (D) Results of automated segmentation of cell membrane. Insets: histogram of best JI distributions. Individual data points outside of the plot range: (C) trained: 0/140; untrained: 10/140; (D) trained: 6/140; untrained: 18/140. To see this figure in color, go online.