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

Area of Science:

• Cellular Biology
• Biophysics
• Medical Imaging

Background:

• Mitochondrial morphology is crucial for cellular function.
• Quantifying mitochondrial morphology at scale is challenging due to limitations in imaging resolution and segmentation tools.
• Soft X-ray tomography (SXT) offers high-resolution, native-state 3D whole-cell imaging, but organelle segmentation remains a bottleneck.

Purpose of the Study:

• To develop a data- and parameter-efficient 3D deep learning model for segmenting mitochondria and nuclei in SXT tomograms.
• To enable high-throughput quantitative analysis of mitochondrial morphology and biophysical properties.

Main Methods:

• Developed MitoXRNet, a 3D deep learning model utilizing multi-axis 3D slicing and Sobel filter-based boundary enhancement.
• Employed a combined Binary-Cross-Entropy and Robust-Dice loss function for optimized segmentation.
• Validated performance on INS-1E cells and tested generalization on unseen data.

Main Results:

• MitoXRNet achieved a 73.8% Dice score with only 1.4 million parameters, outperforming existing methods.
• A larger 22.6 million parameter variant demonstrated strong generalization capabilities.
• Automated segmentation revealed metabolic stimuli-induced mitochondrial remodeling: glucose increased volume and density, while GIP/GKA increased number and density, indicating smaller, dynamic populations.

Conclusions:

• MitoXRNet provides a scalable and efficient framework for segmenting organelles in native-state SXT data.
• The model facilitates quantitative morphological and biophysical profiling of mitochondria.
• This approach enables deeper insights into cellular function and organelle dynamics under various conditions.