How levalby/miswsssm works
Project type and purpose
• This is a research-oriented machine-learning codebase for 3D medical image segmentation.
• It provides two neural architectures (U-Mamba and Swin-UMamba) built on the nnU-Net pattern, plus dataset-adaptation scripts.
• It is designed to run on an HPC cluster (SLURM, NVIDIA A40 GPUs).
Sketch the high-level components
A. Data-management layer
– Datasets/HarDNet-MSEG
– Datasets/Kvasir-SEG
– Preprocessing scripts (download, convert, adapt)
B. Model implementation layer
– U-Mamba (nnU-Net structure + state-space modules)
– Swin-U-Mamba (Vision Transformer backbone + state-space modules)
C. Training orchestration layer
– SLURM batch scripts
– CUDA drivers, conda/mamba environments
– HPC job scheduler interface (bladecluster)
D. Outputs & evaluation
– Trained model checkpoints
– Validation & test metrics
– Visualization / segmentation masks
Relationships and data flow
• Dataset → Preprocessing scripts → Standardized nnU-Net folder structure
→ “run_training.py” in each model directory.
• run_training.py → Submits SLURM job → Allocates GPU via CUDA driver
• GPU nodes → Execute training loops using PyTorch/TensorFlow under nnU-Net APIs
• Models write checkpoints to shared HPC storage → Post-processing/evaluation
• Logs and results → README-referenced output folders
Architectural patterns & principles
• Modular design: Separate folders for each model variant (U-Mamba vs. Swin-U-Mamba)
• HPC-first deployment: Use of SLURM scripts and environment modules reflects batch-processing pattern.
• nnU-Net’s automated pipeline rules: folder conventions drive data split and augmentation.