What it does

NiftyNet is a modular TensorFlow 1.x platform for building and sharing convolutional neural networks on medical images—think 3D brain scans, not cat photos. It bundles pre-trained implementations of specialist architectures like 3D U-net, V-net, and DeepMedic, plus evaluation metrics tuned for segmentation tasks. The consortium behind it spans King’s College London and UCL medical imaging groups.

The interesting bit

The README leads with a rare act of research-software honesty: a 2020 obituary redirecting users to MONAI, the project’s successor. Before that handoff, NiftyNet tackled the fiddly reality of medical dimensions—2.5-D (slice stacks pretending to be volumes), 3-D, and 4-D multi-modal co-registered volumes—rather than forcing everything into standard 2-D pipelines.

Key highlights

• Modular network components designed for sharing pre-trained models between research groups
• Multi-GPU training support for volumetric data
• Built-in implementations of HighRes3DNet, 3D U-net, V-net, and DeepMedic
• “2.5-D” and 4-D input handling for non-standard medical imaging formats
• Explicitly not for clinical use—research-only, per the disclaimer

Caveats

• Unmaintained since April 2020; pinned to TensorFlow 1.15, which is itself end-of-life
• The team formally recommends migrating to MONAI
• Python 2/3 compatibility suggests legacy baggage; no indication of modern TensorFlow or PyTorch support

Verdict

Worth a look if you’re maintaining legacy medical-imaging pipelines or studying how research platforms evolve. Everyone else should follow the maintainers’ own advice and head to MONAI.