VEMamba: Efficient Isotropic Reconstruction of Volume Electron Microscopy with Axial-Lateral Consistent Mamba

TL;DR

This work introduces a realistic degradation simulation and then leverage Momentum Contrast (MoCo) to integrate this degradation-aware knowledge into the network for superior reconstruction of isotropic reconstruction of volume Electron Microscopy datasets.

Abstract

Volume Electron Microscopy (VEM) is crucial for 3D tissue imaging but often produces anisotropic data with poor axial resolution, hindering visualization and downstream analysis. Existing methods for isotropic reconstruction often suffer from neglecting abundant axial information and employing simple downsampling to simulate anisotropic data. To address these limitations, we propose VEMamba, an efficient framework for isotropic reconstruction. The core of VEMamba is a novel 3D Dependency Reordering paradigm, implemented via two key components: an Axial-Lateral Chunking Selective Scan Module (ALCSSM), which intelligently re-maps complex 3D spatial dependencies (both axial and lateral) into optimized 1D sequences for efficient Mamba-based modeling, explicitly enforcing axial-lateral consistency; and a Dynamic Weights Aggregation Module (DWAM) to adaptively aggregate these reordered sequence outputs for enhanced representational power. Furthermore, we introduce a realistic degradation simulation and then leverage Momentum Contrast (MoCo) to integrate this degradation-aware knowledge into the network for superior reconstruction. Extensive experiments on both simulated and real-world anisotropic VEM datasets demonstrate that VEMamba achieves highly competitive performance across various metrics while maintaining a lower computational footprint. The source code is available on GitHub: https://github.com/I2-Multimedia-Lab/VEMamba

Figures (7)

• Figure 1: At the top is the Deep Learning for Isotropic Reconstruction. Compared with isotropic techniques, anisotropic techniques produce fewer axial pixels when imaging the same tissue. Deep learning is then used to compensate for the missing axial pixels. At the bottom is the Comparison between the common method and VEMamba. The common method takes lateral sections as input and reconstructs a volume by stacking, leading to poor axial consistency. In contrast, VEMamba takes volumetric input and performs multi-directional, multi-dimensional scanning, yielding rich axial information.
• Figure 2: Visualization of VEMamba. At the top is the overall structure of VEMamba, which is primarily composed of the Residual Volume Mamba Group (RVMG), and the RVMG consists mainly of the Residual Volume Mamba Block (RVMB). On the bottom left is the Degradation Extraction. On the bottom middle is the Volume Mamba Module (VMM). On the bottom right is the Volume Degradation Injection Module (VDIM).
• Figure 3: Visualization of the VEMamba Module (VEMM). It includes our proposed Axial-Lateral Chunking Selective Scan Module (ALCSSM), Dynamic Weights Aggregation Module (DWAM) and SSM.
• Figure 4: Qualitative comparison of baseline (Interpolation), IsoVEM, EMDiffuse, VEMamba, and Ground Truth on the EPFL dataset at scale factors of ×4, ×8, and ×10 in the lateral (xy) and axial (xz and yz) section.
• Figure 5: Visual comparison of baseline (Interpolation), IsoVEM, EMDiffuse, VEMamba, and Ground Truth on the CREMI dataset at scale factors of ×8 and ×10 in the lateral (xy) section.