ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading

Researchers released ConnectomeBench2, a unified benchmark dataset containing over 716,000 expert-labeled proofreading decisions for automated 3D brain reconstruction across four species. A Vision Transformer model trained on this dataset achieved human-level accuracy in identifying segmentation errors, advancing the automation of connectomic proofreading—a critical bottleneck in neuroscience research.

ConnectomeBench2 addresses a fundamental challenge in computational neuroscience: the labor-intensive process of correcting segmentation errors in electron microscopy reconstructions of neural tissue. Manual proofreading has constrained the pace of connectomics research, limiting the scale and speed at which researchers can map neural circuits. By releasing a comprehensive multi-species dataset spanning mouse, human, zebrafish, and fruit fly brains, the researchers establish a foundation for developing generalizable automated solutions.

The benchmark's significance extends beyond immediate accuracy metrics. The Vision Transformer achieves human-level performance on split error correction while demonstrating robust calibration—meaning its confidence assessments align with actual accuracy. This calibration property is crucial for practical deployment, as it enables researchers to identify cases where the model should defer to human judgment. The authors demonstrate that performance scales predictably with dataset size and image modality, providing a roadmap for further improvements.

The implications span multiple research communities. Neuroscientists gain access to tools that can accelerate connectome reconstruction timelines substantially. The connectomics-specific pretraining and active learning approaches detailed in the research suggest pathways to extend automation to understudied species and brain regions with minimal additional labeling. For AI researchers, the benchmark validates Vision Transformer architectures for specialized 3D geometric and imaging analysis tasks, opening possibilities for applying similar approaches to other biomedical reconstruction problems.

The public release of both dataset and codebase on Hugging Face and GitHub democratizes access, enabling broader research participation and accelerating the development of next-generation proofreading systems.

• →ConnectomeBench2 contains 716,485 expert-labeled proofreading decisions across four major connectome projects, establishing the largest unified benchmark for automated brain reconstruction.
• →Vision Transformer models trained on the dataset achieve human-level accuracy for split error correction and merge error identification, with well-calibrated confidence predictions.
• →Performance improves predictably with dataset size and imaging modality, suggesting systematic pathways to extend automation to new species and brain regions.
• →Connectomics-specific pretraining and active learning reduce labeling requirements, lowering barriers to automating proofreading in understudied neural systems.
• →Open-source release of dataset and code on Hugging Face and GitHub enables collaborative research advancement across the neuroscience and AI communities.