Re-Isotropic Segmentation for Subcellular Ultrastructure in Anisotropic EM Images

Area of Science:

• * Computational biology and structural analysis
• * Advanced imaging and microscopy techniques

Background:

• * Serial section electron microscopy (EM) yields anisotropic data with lower z-axis resolution, hindering accurate 3D reconstruction and segmentation of subcellular structures.
• * Existing 2.5D convolutional methods struggle to fully leverage anisotropic features due to a lack of isotropic scale priors.

Purpose of the Study:

• * To develop a novel framework, ReIsoSeg, that integrates an isotropic scaling prior into anisotropic biovolumes for improved subcellular ultrastructure segmentation.
• * To enhance the connectivity smoothness and segmentation accuracy in 3D EM datasets.

Main Methods:

• * ReIsoSeg employs an anisotropic primary encoder, a pseudo-isotropic auxiliary module for feature deformation, and a weight-shared decoder.
• * A re-isotropic loss function aligns pseudo-isotropic representations with anisotropic labels.
• * The auxiliary module is deactivated during inference, ensuring efficient application.

Main Results:

• * ReIsoSeg demonstrated high performance across multiple challenging datasets, including AC3/AC4, CREMI, and MitoEM.
• * The framework effectively addresses the limitations of anisotropic data in 3D EM imaging.

Conclusions:

• * The proposed ReIsoSeg framework successfully incorporates an isotropic prior to overcome the inherent anisotropy in EM data.
• * This approach significantly improves the accuracy and reliability of subcellular ultrastructure segmentation.