Semi-Automatic Generation of Tight Binary Masks and Non-Convex Isosurfaces for Quantitative Analysis of 3D Biological Samples

Sourabh Bhide, Ralf Mikut, Maria Leptin, Johannes Stegmaier

Abstract

Quantitative analysis of 3D+t light-sheet microscopy data is often hindered by low signal-to-noise ratios and complex, non-convex biological geometries. This paper presents a semi-automatic pipeline that combines user-guided tight binary masking with a geometry-agnostic "unwrapping" technique. This approach prevents segmentation leakage and preserves structural integrity in complex tissues, such as the Drosophila ventral furrow. Validation on embryonic datasets demonstrates high accuracy in extracting cellular metrics like volume and apical area with significantly reduced manual effort compared to fully manual analysis.

Introduction

In developmental biology, 3D time-lapse (3D+t) light-sheet microscopy is essential for observing live cellular dynamics. However, imaging deep into specimens often results in significant signal degradation due to light diffraction. Standard automated segmentation tools, such as the watershed algorithm, frequently fail in these "noisy" environments, leading to boundary "leaks" into the background or adjacent cells.

Furthermore, biological specimens often possess complex, non-convex shapes—for instance, the inward-folding ventral furrow during Drosophila gastrulation. Traditional cartographic projections typically assume simple spherical or ellipsoidal geometries, which causes massive distortions or data loss in these critical regions. This research addresses these challenges by introducing a suite of semi-automatic, open-source tools designed to bridge the gap between error-prone automation and labor-intensive manual analysis.

Methods

The proposed pipeline utilizes a multi-step approach to ensure high-fidelity data extraction:

  1. Tight Binary Masking: To prevent segmentation algorithms from leaking into the background, users label the inner and outer boundaries of a few representative slices. The software then employs piecewise linear interpolation to generate a tight 3D "safety margin" mask for the entire volume.
  2. Non-Convex Surface Projection: The tool creates a "peel" of the 3D surface using a Euclidean Distance Map (EDM) to identify apical (outer) and basal (inner) surfaces. Unlike methods that assume a fixed geometry, this tool "unwraps" surfaces slice-by-slice, allowing it to faithfully follow non-convex folds.
  3. Peel Rectification: A semi-automatic module allows users to center and straighten features of interest (e.g., the ventral furrow) in the final 2D projection by clicking points along the feature.
  4. Segmentation & Tracking: The resulting 2D projections are segmented via a watershed algorithm. Users can interactively add or remove "seeds" to correct errors. These segments are then tracked over time using nearest-neighbor logic.
The semi-automatic pipeline for tight binary masking and non-convex surface projection.

Results and Key Insights

The tools were validated using 3D+t light-sheet microscopy of Drosophila embryos (GAP43::Venus marker) and high-resolution 2-photon confocal microscopy.

Efficiency and Accuracy

The masking procedure requires approximately 2 minutes per time point, while manual seed correction for 2D segmentation takes 3–5 minutes per slice. This "human-in-the-loop" approach achieves high accuracy with a fraction of the labor required for fully manual labeling. By providing a tight, semi-automatically generated mask, the pipeline effectively "corrals" the watershed algorithm, drastically reducing errors in low-signal areas.

Preservation of Complex Geometries

A key insight of this work is the geometry-agnostic unwrapping method. By tracing boundaries slice-by-slice, the method preserves the integrity of complex tissues like the ventral furrow, which would be distorted by standard cartographic methods. The 3D boundary tracing algorithm was shown to faithfully preserve cell areas in 2D projections.

2D and 3D segmentation results alongside quantitative analysis of cell dynamics.

Quantitative Analysis

The pipeline successfully extracted temporal changes in critical cellular metrics, including cell volume, apical surface area, and apical-basal length. These measurements provide a high-resolution view of the mechanics of gastrulation, as shown in the quantitative plots in Figure 2.

Discussion and Future Work

While the pipeline significantly reduces the manual burden, human intervention remains a bottleneck for very long or high-frequency time-lapse movies. Additionally, the current projection method assumes relative uniformity along the sample's long axis; extreme curvature at the poles of an embryo may still result in underestimated cell areas.

Future developments may involve the integration of Convolutional Neural Networks (CNNs) to automate initial foreground identification. This could potentially eliminate the need for manual labeling of the first and last slices, further streamlining the quantitative analysis of complex 3D biological samples.

Source Code: The tools are open-source and available at github.com/stegmaierj/CellShapeAnalysis/.

Summary updated

Abstract

Current in vivo microscopy allows us detailed spatiotemporal imaging (3D+t) of complete organisms and offers insights into their development on the cellular level. Even though the imaging speed and quality is steadily improving, fully-automated segmentation and analysis methods are often not accurate enough. This is particularly true while imaging large samples (100um - 1mm) and deep inside the specimen. Drosophila embryogenesis, widely used as a developmental paradigm, presents an example for such a challenge, especially where cell outlines need to imaged - a general challenge in other systems as well. To deal with the current bottleneck in analyzing quantitatively the 3D+t light-sheet microscopy images of Drosophila embryos, we developed a collection of semi-automatic open-source tools. The presented methods include a semi-automatic masking procedure, automatic projection of non-convex 3D isosurfaces to 2D representations as well as cell segmentation and tracking.
BibTeX
@inproceedings{bhide2020semi,
  author = {Sourabh Bhide and Ralf Mikut and Maria Leptin and Johannes Stegmaier},
  title = {{Semi-Automatic Generation of Tight Binary Masks and Non-Convex Isosurfaces for Quantitative Analysis of 3D Biological Samples}},
  booktitle = {ICIP},
  year = {2020},
  doi = {10.1109/icip40778.2020.9190951},
}
Figure