Diffeomorphic Transport and Prior-Guided Generation for Isotropic Volume EM Reconstruction

Volume electron microscopy (vEM) provides nanometer-scale imaging of cellular ultrastructure, enabling the study of neural circuits and subcellular morphology at unprecedented details. However, because the axial sampling interval is typically much larger than the in-plane resolution, the acquired volumes are highly anisotropic, which limits downstream 3D reconstruction, segmentation, and morphological analysis. Isotropic vEM reconstruction is therefore of central importance, as it recovers anatomically plausible intermediate sections between observed slices. Here we propose a new framework for isotropic reconstruction. Unlike existing axial super-resolution paradigms, we formulate the problem as a structured combination of geometry-preserving transport and detail-restoring generation. This perspective leads to a two-stage design, in which a continuous-time transport model ensures geometric fidelity of structures, while a compact-prior diffusion process recovers high-frequency details. In addition, our method supports arbitrary continuous interpolation between observed sections. Extensive experiments on four datasets from different cell types show that our approach better preserves ultrastructural texture and axial coherence, while achieving superior accuracy and inference efficiency compared with existing methods.