Benchmark Datasets and Pretrained Models for BRIDGE: Whole-Brain Registration Framework

This repository provides the core benchmark datasets, downstream validation image volumes, and trained deep-learning model checkpoints for the manuscript: "BRIDGE: A Scalable Registration Framework for Terabyte-Scale Multi-Round Whole-Brain Microscopy".

Due to storage and bandwidth constraints associated with terabyte-scale volumetric microscopy, this repository deposits the essential representative subsets required to reproduce the paper's key experimental findings, quantitative benchmarks, and downstream transcriptomic applications.

Package Contents and File Descriptions

• brain_registration_with_atlas.rar (Cross-Subject Brain and CCFv3 Atlas Mapping)

  Contains representative volumetric image volumes and corresponding regional annotations demonstrating cross-subject whole-brain registration and alignment to the Allen Mouse Brain Common Coordinate Framework (CCFv3). Includes paired volumes before and after spatial transformation to facilitate visual and quantitative assessment of morphological alignment.

• Data_Multi_round_GABA.rar (Multi-Round Spatial Transcriptomic FISH Data)

  Contains raw multi-round regions of interest (ROIs) from fluorescence in situ hybridization (FISH) profiling targeting GABAergic interneuron subtypes (Pv, Sst, Vip, Vgat). This dataset supports the downstream validation of functional channel realignment and molecular identity preservation across sequential imaging rounds.

• nuclei_valiation_dataset.rar (Cellular-Resolution Nuclear Benchmark Dataset)

  Comprises 50 randomly sampled 3D sub-volumes with segmented cell somatic instances and ground-truth nuclear landmarks. This benchmark dataset is designed to validate cellular-level spatial registration fidelity, sub-voxel centroid correspondence, and the reported 3.71 μm residual distance.

• SwinB-Net_model_data_loss.zip (Pretrained Weights and Training Logs)

  Includes the official pre-trained model weights (checkpoints) for SwinB-Net across progressive hierarchical scales, alongside training convergence loss curves and optimization log metadata.