Deep learning for brain electron microscopy segmentation: Advances, challenges, and future directions in connectomics and ultrastructure analysis

This systematic review and meta-analysis comprehensively analyzes deep learning approaches for brain electron microscopy (EM) segmentation, addressing the critical challenge of extracting neuroanatomical information at nanometer resolution. Following PRISMA guidelines, we identified 60 studies through structured database searches, with quantitative meta-analysis of 27 studies (46 experiments) across 10 datasets providing the first unified benchmark comparison in this domain. Our analysis reveals a field transitioning from traditional CNN approaches toward foundation models and hybrid architectures. The meta-analysis demonstrates that foundation models outperform traditional CNNs by 13%–35% across key metrics, with the 3D Transformer + U-Net achieving the highest composite score (0.954) across five datasets. Meta-analysis confirms significant advantages for foundation models in instance-based metrics (Cohen's d =−6.44), while only 26% of experiments validate across multiple datasets. Four key evolutionary trends emerge: (1) transition from 2D to 3D architectures optimized for ultrastructural complexity; (2) development of topology-preserving loss functions and evaluation metrics (clDice, ERL) that prioritize neural connectivity over pixel-wise accuracy; (3) emergence of self-supervised and foundation model adaptation techniques reducing annotation dependency; and (4) evolution toward specialized architectures capturing long-range dependencies critical for neural structures. Performance analysis reveals that mitochondria segmentation achieves highest accuracy (Jaccard scores 87.2–90.5%), while computational requirements vary from single-GPU implementations to distributed systems with 48 GPUs for teravoxel-scale volumes. Despite progress, reproducibility challenges persist with only 54% of studies providing public code repositories. These advances drive innovation in 3D computer vision, establish new benchmarks for volumetric instance segmentation, and address fundamental challenges in processing massive biological datasets. Our unified benchmarks and comprehensive analysis provide a foundation for systematic progress tracking and evidence-based method selection, positioning brain EM segmentation to enable large-scale connectomics studies and detailed neuroanatomical mapping across scales.