Automated brightfield morphometry of 3D organoid populations by OrganoSeg

Abstract

Spheroid and organoid cultures are powerful in vitro models for biology, but size and shape diversity within the culture is largely ignored. To streamline morphometric profiling, we developed OrganoSeg, an open-source software that integrates segmentation, filtering, and analysis for archived brightfield images of 3D culture. OrganoSeg is more accurate and flexible than existing platforms, and we illustrate its potential by stratifying 5167 breast-cancer spheroid and 5743 colon and colorectal-cancer organoid morphologies. Organoid transcripts grouped by morphometric signature heterogeneity were enriched for biological processes not prominent in the original RNA sequencing data. OrganoSeg enables complete, objective quantification of brightfield phenotypes, which may give insight into the molecular and multicellular mechanisms of organoid regulation.

Figure 1

The OrganoSeg pipeline segments brightfield 3D-culture images more accurately than existing methods. (a) Key steps in the image-segmentation pipeline applied to a representative brightfield image. Users can opt to split aggregates (blue) or remove debris (red) in the software. (b,c) Comparison of OrganoSeg with competing alternatives^– according to spheroid call rates in (b) and Kolmogorov-Smirnov statistics (K-S stat) of segmented area distributions in (c). The number of images segmentable by each algorithm is shown to the right of c. Images were segmented with following OrganoSeg parameters: Otsu threshold = 1, Max-window size = 250 pixels, Size-exclusion threshold = 10 pixels. Data in (b) are shown as the median call rate with 95% confidence intervals in brackets from n = 19 images of MCF10A-5E spheroids. Data in (c) are shown as median boxplots of K-S stat from 1000 bootstrap replicates of n = 861 total spheroids, with significant differences from manual segmentation (gray) assessed by K-S test.

Figure 2

Data-driven classification and quantification of spheroid states by OrganoSeg. (a–c) Spheroid polydispersity is associated with an enlarged state in a fraction of triple-negative breast cancer lines. (d–f) A discrete round state is specifically enriched in nontransformed MCF10A-5E 3D cultures. (g–i) Reproducible differences in stellate invasion among triple-negative breast cancer lines. tSNE plots in (a,d,g) are pseudocolored by the weight of each image metric in the projection, and gates for the indicated spheroid states (gray) are dashed. Projections of the indicated cell lines in (b,e,h) are shown alongside specific brightfield examples (stars). Images were segmented with following OrganoSeg parameter ranges: Otsu threshold = 0.183–1, Max-window size = 100–370 pixels, Size-exclusion threshold = 148–1130 pixels. Data in (c,f,i) are shown as the mean ± s.d. of n = 4 cultures together containing 102–1069 spheroids.

Figure 3

OrganoSeg identifies quantitative heterogeneity and commonality among colorectal cancer organoids. (a) Hierarchical clustering of tSNE projections (Supplementary Fig. S5) binned for 10–434 organoids in 64 subclones across three patients. The clustering position of specific image subpanels is indicated. (b–d) Subclones within patients exhibit substantially different size distributions and optical-density features. (e–g) Examples of subclones in other patients with binned tSNE projections similar to those in (b) to (d). The insets of (b) to (g) show the projection of the corresponding subclone on the tSNE plot of all 4940 tumor organoids in the dataset. Images were segmented with following OrganoSeg parameter ranges: Otsu threshold = 0.147–1, Max-window size = 30–260 pixels, Size-exclusion threshold = 50–2000 pixels.

Figure 4

Morphometric profiling fuses colorectal cancer organoids that patient-segregate by RNA sequencing. (a) Hierarchical clustering of RNA sequencing data for the indicated patient (P), tumor fragment (T), and subclone. Enriched Gene Ontology biological processes for the indicated gene clusters are included in Supplementary Tables S3–S7. (b,c) Hierarchical clustering of morphometric profiles based on median OrganoSeg metrics in (b) and interquartile range in (c).

Figure 5

mRNA-morphometry associations suggest new hypotheses for organoid regulation. (a,b) Hierarchical clustering of Pearson correlations between transcript abundance and median OrganoSeg metrics in (a) and interquartile range in (b). Enriched Gene Ontology biological processes for the indicated gene clusters are included in Supplementary Tables S8–S10. Braces in (a) indicate gene-set enrichment for Sp1/Krüppel-like factor (KLF) binding sites in the cluster.