Robust mitochondria segmentation and morphological profiling using soft X-ray tomography

Abstract

Mitochondrial morphology is central to cellular function, yet large-scale quantification is limited by the lack of high-resolution whole-cell imaging and efficient segmentation tools. Soft X-ray tomography (SXT) provides native-state 3D whole-cells images, but organelle segmentation remains a bottleneck. We present MitoXRNet, a data- and parameter-efficient 3D deep learning model for mitochondria and nucleus segmentation in SXT tomograms. Using multi-axis 3D slicing, Sobel filter-based boundary enhancement, and a combined Binary-Cross-Entropy and Robust-Dice loss, MitoXRNet achieves a 73.8% Dice score on INS-1E cells with only 1.4 M parameters, outperforming existing methods. A larger 22.6 M variant generalized well to unseen data. Automated segmentation enabled quantitative analysis of mitochondrial remodeling under metabolic stimuli: glucose increased mitochondrial volume and matrix density, while GIP and GKA increased mitochondria number, reduced volume, and elevated density, indicating smaller, denser, more dynamic populations. MitoXRNet provides a scalable framework for high-throughput morphological and biophysical profiling of organelles in native-state SXT data.

Keywords: Deep Learning; Gastric Inhibitory Polypeptide (GIP); Glucokinase Activator (GKA); Mitochondria Remodeling; Mitochondria Segmentation; Soft X-Ray Tomography.