Latent-centric Isotropic Resolution Enhancement for Expansion Microscopy Imaging via Neural Compression and Self-supervised Learning
Expansion microscopy (ExM) enables nanoscale imaging for disease characterization. However, whole-organ analyses remain limited by several challenges. Current super-resolution methods either require high-resolution ground-truth data or assume spatially uniform point spread functions--assumptions that rarely hold in whole-organ imaging with depth-varying aberrations and illumination drift. Existing methods also worsen storage demands by inflating already multi-terabyte datasets without using neural compression. We propose a single-stage, self-supervised framework that addresses both resolution anisotropy and storage constraints through compression-aware isotropic super-resolution. Our approach combines a 2D lateral encoder that operates directly on raw slices to avoid memory limits with a lightweight volumetric decoder that preserves cross-slice continuity. A vector-quantized variational autoencoder (VQ-VAE) provides an information-sufficient bottleneck, achieving up to 128x slice compression and up to 8x axial resolution enhancement. This latent-centric design yields approximately 1000x reduction in storage compared with storing fully isotropic volumes. The framework achieves higher GPU throughput, lower memory usage, and stronger multi-GPU scalability than prior methods. By designating compressed latent space as the native storage format, it enables efficient on-demand isotropic reconstruction directly from compact representations. This combination of isotropic enhancement and neural compression framework therefore makes large-scale, whole-organ ExM analysis practical while maintaining analysis-ready accessibility, addressing a bottleneck in translating ExM to clinical biomarker discovery.