ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading

ConnectomeBench2 is a newly released benchmark dataset consisting of over 716,485 expert-labeled proofreading decisions and more than 4.5 million associated images from major open connectomes across multiple species, including mouse, human, zebrafish, and fly. A Vision Transformer model trained on this dataset achieves human-level accuracy for split and merge error corrections, demonstrating scalability with data size and modality. This resource is significant for researchers in connectomics as it offers a unified framework for training and evaluating advanced vision models, potentially reducing labeling efforts for new species and brain regions through its connectomics-specific pretraining and active learning strategies.