Medical Segmentation’s Abandoned Model Zoo

What it does

This repository collects implementations of popular medical image segmentation networks—U-Net, U-Net++, VNet, UNETR, and others—into a single PyTorch codebase for both 2D and 3D volumes. It handles various medical formats like nii.gz and mhd through configuration in hparam.py and processes images in patches so they do not need identical dimensions.

The interesting bit

The project is largely curation and glue: it borrows heavily from MedicalZooPytorch and torchio, and several models are copied directly from other authors’ repositories. The patch-based pipeline is a practical concession to the reality that medical scans rarely share the same resolution.

Key highlights

• Supports 2D and 3D variants of eight segmentation architectures each, including DeepLab, PSPNet, and DenseVoxelNet
• Compatible with common medical imaging formats via fold_arch configuration
• Patch-based training and inference avoids rigid size requirements
• Includes a metrics module for evaluation
• Authors have explicitly deprecated the repository in favor of SLab-Medical-Segmentation

Caveats

• The repository is deprecated; the authors direct users to a successor project
• Several network implementations are copied from other repositories rather than reimplemented
• Multi-category segmentation requires manual code modification, and the authors caution that the project “is not perfect and there are still many problems”

Verdict

Useful if you need a quick survey of how standard medical segmentation models look in PyTorch, but start with the authors’ successor repository if you want active development.