MedNeXt

Why we included this project

Clinical CT and MRI segmentation often runs into a labeling bottleneck. MedNeXt, a fully convolutional 3D architecture from the German Cancer Research Center, was built for those sparse-annotation settings. It reuses the ConvNeXt block design and scales it to volumetric data, so you get a strong baseline for multi-class segmentation without needing a pile of annotated scans. The code ships as PyTorch modules, which lets you drop in the full network or just individual blocks and keep your own training loop, or use the bundled nnU-Net-based pipeline from the MICCAI 2023 paper. Model variants from small to large and gradient checkpointing help fit GPU-limited setups. Good fit for researchers and small medical imaging teams that want a reproducible starting point.