niiv: Interactive Self-supervised Neural Implicit Isotropic Volume Reconstruction

niiv: Interactive Self-supervised Neural Implicit Isotropic Volume Reconstruction

Troidl J, Liang Y, Beyer J, Tavakoli M, Danzl JG, Hadwiger M, Pfister H, and Tompkin J.

MICCAI Workshop on Efficient Medical AI (EMA), 2025.

Three-dimensional (3D) microscopy data often is anisotropic with significantly lower resolution (up to 8x) along the z axis than along the xy axes. Computationally generating plausible isotropic resolution from anisotropic imaging data would benefit the visual analysis of large-scale volumes. This paper proposes niiv, a self-supervised method for isotropic reconstruction of 3D microscopy data that can quickly produce images at arbitrary output resolutions. The representation embeds a learned latent code within a neural field that describes the implicit higher-resolution isotropic image region. We use a novel attention-guided latent interpolation approach, which allows flexible information exchange over a local latent neighborhood. Under isotropic volume assumptions, we self-supervise this representation on low-/high-resolution lateral image pairs to reconstruct an isotropic volume from low-resolution axial images. We evaluate our method on simulated and real anisotropic electron (EM) and light microscopy (LM) data. Compared to a state-of-the-art diffusion-based method, niiv shows improved reconstruction quality (+1dB PSNR) and is over three orders of magnitude faster (1,000x) to infer. Specifically, niiv reconstructs a 128^3 voxel volume in 2/10th of a second, renderable at varying (continuous) high resolutions for display.

Acknowledgements

This work was supported by NIH grants 1U01NS132158 and R01HD104969. We thank the reviewers for their constructive feedback.