DeepBranchAI: a transferable 3D segmentation model for branching networks

Three-dimensional branching networks exist throughout biological, natural, and engineered systems as pathways through volumetric space. Segmentation is required to reconstruct these networks for analysis, but minor voxel misclassifications can alter connectivity by falsely disconnecting continuous structures or amplifying spurious branches. Addressing this topological vulnerability requires true 3D segmentation models, since 2D slice-by-slice approaches cannot maintain connectivity across x, y, and z axes. Here, we present DeepBranchAI, a 3D nnU-Net model trained on expert-refined mitochondrial FIB-SEM reference labels as a transferable checkpoint for topology-sensitive segmentation of branching networks. Skeletal-muscle mitochondrial reticulum provides a dense and heterogeneous branching-network training target, and FIB-SEM captures this structure with isotropic nanometer-scale detail. Training on this dataset produced a 3D checkpoint with DSC = 0.942 ± 0.020 across five-fold mitochondrial FIB-SEM validation, providing the basis for subsequent transfer-learning experiments. Transfer learning was evaluated with five-fold cross-validation on paired image/reference-label datasets from 3D-IRCADb venous segmentation, Plant CT roots, and AeroPath airway CT, using continuity-sensitive measures including clDice and absolute connected-component error. Across the external datasets, DeepBranchAI-pretrained fine-tuning had higher mean clDice than scratch nnU-Net. Relative differences in mean clDice were 11.7% for 3D-IRCADb, 5.5% for Plant CT roots, and 1.9% for AeroPath. Mean absolute connected-component error was 25.0% lower for Plant CT roots and 11.7% lower for AeroPath. In 3D-IRCADb, DeepBranchAI-pretrained fine-tuning also had higher mean Dice and clDice than VesselFM, 0.679 versus 0.464 and 0.629 versus 0.360, respectively. The transfer results suggest that training on structurally heterogeneous skeletal-muscle mitochondrial networks can produce 3D features useful beyond the original FIB-SEM domain. Code, trained weights, external-transfer scripts, and validation tables are provided as open-source resources.