PyTorch Connectomics

Built on PyTorch Lightning and MONAI, enabling automatic distributed training, mixed precision, and medical imaging optimizations without manual configuration.

Includes MONAI models like UNETR and MedNeXt architectures, which are benchmarked for accurate 3D segmentation in connectomics, as shown in the tutorials with high Jaccard indices.

Features Waterz decoder with hierarchical agglomeration and dust merge for fragment cleanup, essential for refining EM segmentations with tunable parameters via Optuna.

Supports multi-GPU training, efficient data loading with caching, and gradient accumulation, allowing it to handle large-scale datasets from single GPU to clusters.