MitoStructSeg

Mitochondrial morphology is a direct readout of cellular energy metabolism, and damage to the cristae — the densely folded invaginations of the inner membrane — accompanies cardiomyopathy, ischemic injury, and metabolic disease. Focused ion beam scanning electron microscopy (FIB-SEM) resolves these structures at nanometer scale, but analysis has lagged the imaging: most segmentation tools delineate whole mitochondria as filled blocks and never recover the internal membrane architecture, and those that do address cristae typically need human-in-the-loop correction or fresh annotation per sample.

MitoStructSeg, developed at Beijing Institute of Technology with the cryo-electron microscopy center at Southern University of Science and Technology, couples membrane-and-cristae segmentation to downstream morphometry in one automated framework. Its segmentation engine, AMM-Seg (Adaptive Multi-domain Mitochondrial segmentation), is trained with domain adaptation so a model fit on one annotated sample transfers to unlabeled volumes from other patients, and from human myocardium to mouse kidney, without new manual labeling. A second module converts the masks into per-organelle measurements: outer-membrane surface area, inner-cristae surface area, solid volume, total volume, and cristae density.

The bioRxiv preprint was first posted in June 2024 and revised through July 2025; it has not been peer reviewed. Where general-purpose electron microscopy models such as OmniEM span many organelles and tasks, and annotation-free volumetric tools such as CellSeg3D target fluorescence, MitoStructSeg trades breadth for the depth needed to resolve cristae.

#Key Features

• Cristae-resolved segmentation: Recovers the outer membrane and the inner cristae as distinct thin, line-like structures rather than a single filled organelle mask, enabling ultrastructural rather than merely positional analysis.
• Dual-channel feature fusion: Each 2D slice becomes an augmented image (geometric augmentation on the source domain, intensity augmentation on the target) and a texture image from Gaussian blur plus Canny edge extraction, with channel attention adaptively weighting the two streams.
• Domain adaptation across samples: A gradient reversal layer and domain classifier drive domain-invariant features while pseudo-labels supervise unlabeled target volumes, so a trained model generalizes to new patients and tissues from limited annotation.
• Inter-slice continuity learning: A Cross-Slice Fusion module merges adjacent-slice encodings and a second decoder predicts the XOR difference between neighboring labels, keeping segmentations coherent along z for 3D reconstruction.
• Morphometry and GUI: A quantitative module extracts surface areas, volumes, and cristae density behind a React interface, with 3D Slicer rendering of healthy versus damaged organelles.

#Technical Details

AMM-Seg is a weight-sharing encoder-decoder network with a dual decoder: one head produces segmentation, the other captures inter-slice variation. Training combines a source-domain cross-entropy loss on labeled slices and their inter-slice difference, a pseudo-label loss on the unlabeled target domain, and a domain classification loss behind a gradient reversal layer, the latter two weighted at 0.1. Optimization uses Adam, an initial learning rate of 0.0005 under polynomial decay with power 0.9, batch size 2, and 10,000 iterations on an NVIDIA A100.

The primary dataset is FIB-SEM imaging of myocardial biopsies from three post-COVID-19 patients at 5 x 5 nm pixel size with 10 nm milling steps, in volumes up to 6500 x 5200 x 800 voxels. Two hundred manually annotated 800 x 800 slices from Patient#1 form the labeled source domain, unlabeled 200-slice blocks from each patient serve as target domains, and 40 slices are held out for validation; a public Janelia mouse kidney volume at 8 nm isotropic resolution provides the cross-species test. Against MEDIAR, DA-ISC, U-Net, Swin UNETR, and UNETR, AMM-Seg achieves the highest median F1 with the smallest spread across all three patient datasets and the best IoU, MCC, and F1 on mouse kidney — for full structure, outer-membrane-only, and cristae-only segmentation alike. Absolute scores fall below conventional block segmentation because thin membrane contours are far more sensitive to small errors.

#Applications

The framework targets researchers using volume electron microscopy to study mitochondrial dysfunction — cardiac pathology, nephrology, neurodegeneration — where the phenotype of interest is cristae architecture rather than organelle count. In an analysis of 240 mitochondria across three patients, cristae density separated healthy from damaged organelles far more cleanly than surface-area or volume-occupancy ratios, offering a concrete morphometric endpoint for tissue-level studies. Pretrained checkpoints, inference scripts, and the GUI let imaging groups quantify their own volumes without writing training code.

#Impact

MitoStructSeg addresses a gap that limits how volume electron microscopy data are used: segmentation and quantification are usually separate, manual, sample-specific steps. By making cross-sample transfer the design goal, it offers a path to comparing mitochondrial ultrastructure across patients and cohorts on a consistent basis. The project is early — the preprint awaits peer review, the repository has seen little community uptake, checkpoints and datasets are distributed through cloud-drive links rather than a model hub, and no code license has been attached — and the authors note that generalizing beyond the two validated tissue types would benefit from larger-scale unlabeled pretraining. The clearest contribution is methodological: domain adaptation plus explicit inter-slice continuity suffices to segment cristae from a few hundred annotated slices.