Code provenance and third-party notices
The standalone model consolidates project SlimFormer components: window
partition, MLP/LayerScale, dense window Mamba, aligned MedNeXt decoding,
plans-aware downsampling, refined-trunk routing and subpixel upsampling.
AugmentedMixer is the project-owner-supplied R5 CapFree implementation.
SelectiveMixer is the descriptive name for DenseWindowMamba with the same
computation. Encoder2/3 retain MLP residuals but no token-mixer residuals.
The release model file matches the currently inspected training source.
The Mamba-free implementation (nnUNetTrainer_SlimFormerPlusPlus_MambaFree.py) is a
copy of that file in which SelectiveMixer is replaced by window self-attention
using PyTorch's scaled_dot_product_attention; it uses neither Mamba nor
causal-conv1d.
The source checkout's Apache-2.0 license text, including its attribution to
Division of Medical Image Computing, German Cancer Research Center (DKFZ),
Heidelberg, Germany (2019), is preserved verbatim in
licenses/Apache-2.0.txt. The project research-only license does not replace
third-party licenses or withdraw permissions granted by upstream owners.
Original project contributions: Copyright (c) 2026 WHHHHHY822.
External implementations remain dependencies; their source is not vendored:
- PyTorch
- MONAI
- einops
- Mamba
- causal-conv1d
- MedNeXt, source revision
0b78ed869fbd1cc2fd38754d2f8519f1b72d43ba - nnU-Net
- dynamic-network-architectures
Each dependency retains its own license and notices. Consult the exact installed revision for its terms. No dataset or pretrained weights are included in this source release.