From Instance Segmentation to 3D Growth Trajectory Reconstruction in Planktonic Foraminifera

TL;DR

This paper tackles automated reconstruction of 3D growth trajectories in planktonic foraminifera by integrating instance segmentation with a chamber-ordering step on high-resolution CT data. It introduces an end-to-end pipeline with three segmentation variants (U-Net + SW, 3D PlantSeg, 3D MTL + SW) and a nearest-neighbor chamber-ordering algorithm, validated on a 50-specimen Menardella CT dataset. Across metrics $IoU$, $ARI$, $VI$, and $ ho$, the study shows strong ordering robustness (median $ ho \geq 0.87$) and accurate centroid localization (mean $\delta < 4$ voxels; $3D$ MTL + SW achieving $<1$ voxel). The work demonstrates that automated, reproducible chamber growth analysis is feasible and scalable, with potential to enable large-scale data-driven ecological studies of developmental morphology in chamber-secreting organisms.

Abstract

Planktonic foraminifera, marine protists characterized by their intricate chambered shells, serve as valuable indicators of past and present environmental conditions. Understanding their chamber growth trajectory provides crucial insights into organismal development and ecological adaptation under changing environments. However, automated tracing of chamber growth from imaging data remains largely unexplored, with existing approaches relying heavily on manual segmentation of each chamber, which is time-consuming and subjective. In this study, we propose an end-to-end pipeline that integrates instance segmentation, a computer vision technique not extensively explored in foraminifera, with a dedicated chamber ordering algorithm to automatically reconstruct three-dimensional growth trajectories from high-resolution computed tomography scans. We quantitatively and qualitatively evaluate multiple instance segmentation methods, each optimized for distinct spatial features of the chambers, and examine their downstream influence on growth-order reconstruction accuracy. Experimental results on expert-annotated datasets demonstrate that the proposed pipeline substantially reduces manual effort while maintaining biologically meaningful accuracy. Although segmentation models exhibit under-segmentation in smaller chambers due to reduced voxel fidelity and subtle inter-chamber connectivity, the chamber-ordering algorithm remains robust, achieving consistent reconstruction of developmental trajectories even under partial segmentation. This work provides the first fully automated and reproducible pipeline for digital foraminiferal growth analysis, establishing a foundation for large-scale, data-driven ecological studies.

Figures (7)

• Figure 1: An overview of the developed pipeline for determining chamber growth trajectories in planktonic foraminifera. The process starts with raw CT scan images, which pass through a deep neural architecture like U-Net to produce binary masks that identify chamber regions. A grouping step using hand-crafted methods follows to separate individual chambers, resulting in instance masks. Spatial features extracted from these masks support the inference of the 3D chamber growth trajectory.
• Figure 2: Example images of the four different species of planktonic foraminifera woodhouse2023paleoecology used in our dataset. (a) Menardella exilis, (b) Menardella limbata, (c) Menardella menardii and (d) Menardella pertenuis. The scalar bar for images is 200 microns.
• Figure 3: Statistical distributions of chamber characteristics in our dataset. (a) Histogram of chamber counts per specimen (total $50$ specimens) showing a unimodal distribution (range: 13--24 chambers). (b) Logarithmic distribution of individual chamber sizes (total $1024$ measurements) revealing heavy-tailed morphology spanning $10^2$--$10^6$ voxels, indicating right-skewed size variation.
• Figure 4: Evaluation of predicted instance segmentation compared to ground truth on the test set, along with representative slices of instance segmentation results for four species of Menardella. (a) shows the segmentation accuracy of different chamber segmentation methods on the test set. While the 2D U-Net + SW achieved the highest performance in overall chamber segmentation based on IoU computed on binary masks, the 3D MTL + SW showed comparable performance in instance-level segmentation when evaluated using the ARI on instance masks. (b) shows the segmentation errors, where all segmentation methods tend to under-segment chambers. (c) presents example instance segmentation results by different methods for each species. The first column shows raw CT image slices, the second column shows the ground truth chamber segmentation, and the remaining columns display predictions from the tested methods. Different chamber instances were randomly assigned different colors, while white bounding boxes highlight regions where the methods made errors due to over- or under-segmentation or missing chambers.
• Figure 5: Regression analysis of predictions versus ground truth for measuring the number of chambers on the test set. The scatter plot with jitter shows the relationship between predicted values and actual measurements. Each line represents the fitted linear regression, and the shaded area indicates the 95% confidence interval around the regression line.