EMGauss: Continuous Slice-to-3D Reconstruction via Dynamic Gaussian Modeling in Volume Electron Microscopy

TL;DR

EMGauss tackles the challenge of anisotropic volume electron microscopy by reframing slice-to-3D reconstruction as continuous 3D rendering with deformable Gaussian splats. It introduces a canonical Gaussian representation of the volume, a temporal axial deformation network, and an EMA teacher with pseudo-labeling to achieve high-fidelity, continuous depth interpolation without large external datasets. The approach outperforms diffusion- and GAN-based baselines on both simulated and real anisotropic vEM data, delivering smoother geometry, fewer artifacts, and arbitrary-depth synthesis. This work offers a generalizable, data-efficient framework for slice-to-volume reconstruction that could extend to other planar-imaging modalities beyond vEM.

Abstract

Volume electron microscopy (vEM) enables nanoscale 3D imaging of biological structures but remains constrained by acquisition trade-offs, leading to anisotropic volumes with limited axial resolution. Existing deep learning methods seek to restore isotropy by leveraging lateral priors, yet their assumptions break down for morphologically anisotropic structures. We present EMGauss, a general framework for 3D reconstruction from planar scanned 2D slices with applications in vEM, which circumvents the inherent limitations of isotropy-based approaches. Our key innovation is to reframe slice-to-3D reconstruction as a 3D dynamic scene rendering problem based on Gaussian splatting, where the progression of axial slices is modeled as the temporal evolution of 2D Gaussian point clouds. To enhance fidelity in data-sparse regimes, we incorporate a Teacher-Student bootstrapping mechanism that uses high-confidence predictions on unobserved slices as pseudo-supervisory signals. Compared with diffusion- and GAN-based reconstruction methods, EMGauss substantially improves interpolation quality, enables continuous slice synthesis, and eliminates the need for large-scale pretraining. Beyond vEM, it potentially provides a generalizable slice-to-3D solution across diverse imaging domains.

Figures (6)

• Figure 1: Different paradigms of isotropic reconstruction from anisotropic datasets. Existing approaches typically extract volume features from abundant xy-plane images. After stacking anisotropic slices along the z-axis (top-left), they: (a) Video interpolation along the z-axis: models are trained along the x or y direction, where adjacent frames are input to synthesize the intermediate frame; during inference, the trained model is repurposed to interpolate along the z-axis, generating intermediate slices corresponding to continuous temporal coordinates ($\tau\in[0,1]$). (b) Image super-resolution on xz/yz planes: Since the xy-plane images are of much higher resolution, a low-resolution counterpart can be generated either by manual downsampling emdiffuse or by learning a degradation model cycleganir, forming paired HR–LR supervision for training; during inference, the resulting super-resolution model is then applied along the z-axis to enhance xz/yz slices. Both paradigms implicitly assume spatial isotropy. In contrast, our method directly performs inference in continuous 3D space, without relying on such isotropic assumptions.
• Figure 2: Overview of the pipeline of EMGauss. Given a volume EM dataset (top), the gray-shaded regions denote the missing intermediate slices to be interpolated, while the visible regions represent the available anisotropic training data. We interpret the low-resolution $z$-axis as a temporal dimension and employ a deformable 2D Gaussian splatting representation to model the geometric evolution of 2D structures along depth. During inference, the model synthesizes the target slice by directly conditioning on the relative temporal coordinate corresponding to the desired interpolation position.
• Figure 3: Isotropic xy-slice reconstruction results on EPFL dataset lucchi2013learning.
• Figure 4: Isotropic xy-slice reconstruction results on FIB-25 dataset takemura2015synaptic.
• Figure 5: Isotropic xz/yz-slice Reconstruction results on synthetic anisotropic datasets.