MitoXRNet: Deep Learning-enabled 3D Segmentation of Mitochondria and Nucleus in Soft X-ray Tomograms

Mitochondrial morphology and distribution are central to cellular function, but large-scale quantification is limited by the lack of efficient segmentation tools. Soft X-ray tomography (SXT) enables whole-cell, high-resolution imaging of cells in their native state, but segmentation of organelles remains a major bottleneck. We present MitoXRNet, an automated 3D segmentation method for mitochondria and the nucleus in SXT tomograms. The training data were generated by slicing tomograms into overlapping 3D volumes along all three axes, and a Sobel-based preprocessing step enhancing mitochondrial boundaries. MitoXRNet was trained using Binary Cross-Entropy or a combined BCE with Robust Dice loss to address class imbalance. On INS-1E cell line datasets, a compact model (1.4M parameters) outperformed prior 2D U-Nets and the self-configuring nnU-Net (31.2M) with a Dice score of 73.8%. A larger variant (22.6M) achieved strong generalization across unseen datasets. MitoXRNet enables scalable, accurate, and high-throughput segmentation for systematic studies of mitochondrial morphology and cellular architecture.