deltaMic: Inverse 3D Microscopy Rendering for Embryo Shape Inference with Active Mesh

Inverse 3D Microscopy Rendering for Embryo Shape Inference with Active Mesh

Highlight

Exhibit Hall I #217

Sacha Ichbiah¹
Anshuman Sinha^1,2
Fabrice Delbary¹
Hervé Turlier¹

¹ Collège de France, CNRS, INSERM, PSL University
² Université Paris Cité

Overview

deltaMic is a CUDA-based differentiable 3D renderer for fluorescence microscopy. It models 3D image formation/aquisition as a convolution between the microscope's point-spread function (PSF) and a triangular 3D mesh representation of the specimen. By jointly optimizing mesh geometry and optical parameters directly from raw images, deltaMic achieves robust 3D reconstruction without labeled training data or priors.

Figure 1 - deltaMic pipeline overview.

To keep this webpage results-oriented, we urge the interested reader to explore the neat methodology in greater detail from our full paper!

First Experiment Results

The algorithmic objective of deltaMic is: Given a practical initial 3D mesh and a target 3D image, it will differentiably optimize this mesh to represent the object embedded in the 3D image - thus aquiring a differentiable 3D mesh model of the object!

The following videos show the progression of deltaMic starting from simple meshes — sphere or toroid — to match target geometries in clean, artificial 3D fluorescence images. These demonstrate the optimization's convergence from abstract shapes to realistic targets using synthetic general-purpose data.

A sphere dividing in two

Toroid to a bath Duck

Sphere to a Cow

Sphere to a Fish

Microscopy Reconstruction Results (C. elegans)

This next set of experiments applies deltaMic to real 3D microscopy data, showing its ability to fit fluorescent membrane images of C. elegans embryos at various developmental stages — 4-cell, 8-cell, and 16-cell stages — demonstrating accurate inference of multicellular morphology from real 3D imaging data.

C. elegans - 4-cell stage reconstruction

C. elegans - 8-cell stage reconstruction

C. elegans - 16-cell stage reconstruction

Figure 2 - Optimization resaults on C. elegans embryos across different cell stages.

Results Across Species

deltaMic accurately reconstructs early-stage embryo morphologies across species and imaging modalities.

Figure 3 - Inferred cellular 3D geometries for mouse, ascidian and C. elegans embryos.

Benchmarking

To assess generality, deltaMic was compared against DM3D, a leading active-mesh segmentation framework implemented as a Fiji plug-in.

Figure 4 - Comparison between DM3D and deltaMic on 3D mouse organoid data.

deltaMic demonstrates improved reconstruction of multicellular membrane structures while maintaining photometric fidelity. Performance on nucleus-only datasets is limited due to design assumptions targeting multicellular surfaces.

Discussion

deltaMic bridges physics-based rendering and inverse modeling in biological imaging. This approach enables quantitative analysis of morphogenesis by linking observed fluorescence to underlying 3D geometry and optical parameters.

Current limitations

Dependence on an adequate initial mesh

Ongoing directions

Automated mesh initialization
Temporal morphodynamic inference (tension, curvature, pressure)
Integration with biophysical simulation pipelines

Citation

@InProceedings{deltamic_2025_ICCV,
    author    = {Sacha Ichbiah and Anshuman Sinha and Fabrice Delbary and Hervé Turlier},
    title     = {Inverse 3D Microscopy Rendering for Cell Shape Inference with Active Mesh},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {27466-27475}
}

Inverse 3D Microscopy Rendering for Embryo Shape Inference with Active Mesh

Paper

Code

Video