How bowang-lab/medsam works
Project type and purpose
• MedSAM is a full-stack research toolkit for 2D/3D medical image segmentation.
• Purpose: fine-tune and extend the “Segment Anything” (SAM) model for medical images, provide training pipelines, inference scripts, GUI, and comparison benchmarks.
File structure
• Top-level directories and files:
• segment_anything/ – the core SAM model implementation (encapsulates model architecture, prompt encoder, mask decoder, utilities).
• extensions/ – optional prompt-based segmentation extensions (point, text, sparse-marker).
• comparisons/ – inference and training scripts to benchmark against other methods (DeepLabV3+, nnU-Net, vanilla SAM).
• utils/ – shared utilities for preprocessing, checkpoint conversion, evaluation metrics (SurfaceDice), data formatting.
• train_one_gpu.py, train_multi_gpus.py & train_multi_gpus.sh – training entrypoints and SLURM script for distributed training.
• pre_CT_MR.py – data preprocessing: windowing, normalization, resampling.
• MedSAM_Inference.py, gui.py – inference entrypoints: CLI and PyQt5 GUI.
• work_dir/ – recommended filesystem layout for checkpoints and intermediate data.
README for additional insights
• Architecture: based on Meta’s SAM (Vision Transformer + prompt encoder + mask decoder).
• Dependencies: PyTorch 2.0, cc3d, PyQt5.
• Workflows: data preprocessing → training/fine-tuning → ckpt conversion → inference (CLI, notebook, GUI).
Detail
• Subcomponents within segment_anything (prompt encoder, transformer, mask decoder).
• Optional flows through extensions (e.g., point_prompt/train_point_prompt.py).
• Both single-GPU and distributed-GPU training paths.
• Don’t overload: group similar scripts under “Training Pipeline” but call out the SLURM script separately.