Robust Deep Learning-based 3D Segmentation and Morphological Analysis of Mitochondria using Soft X-ray Tomography

Mitochondrial morphology is crucial for cellular function, but large-scale analysis is limited by challenges in high-resolution imaging and segmentation. MitoXRNet, a compact 3D deep-learning model, efficiently segments mitochondria and nuclei from Soft X-ray Tomography data using multi-axis slicing, Sobel-based boundary enhancement, and combined BCE–Robust Dice loss. With 1.4M parameters, it achieves a Dice score of 73.8% on INS-1E cells, outperforming existing models. Automated analysis indicated that glucose induced larger mitochondria and higher matrix density, and that GIP and GKA induced smaller and denser mitochondria, highlighting previously unreported β-cell mitochondrial remodeling. MitoXRNet allows for scalable profiling of organelle-level morpho-biophysical data. A data-efficient method for 3D segmentation of mitochondria and nucleus from Soft X-ray tomograms. Incorporates domain-specific Sobel filter-based preprocessing to improve segmentation accuracy and quality under imperfect or noisy labels. Enables rapid and automated analysis of mitochondrial morphology, facilitating quantitative assessment of pharmacological effects on cellular ultrastructure.