How zhangyh0502/pcasam works
Project type and purpose
• PCaSAM is a research-oriented, deep-learning segmentation toolkit specialized for prostate cancer on multi-modal MRI. It integrates a medical “foundation” model (Segment Anything Model, SAM) with custom networks (e.g., UNet fusion) to perform training and inference.
Main components of the system
a. Data Ingestion & Preprocessing
• datamodule/ (MRI_CancerSeg, picai_decomp, prostate158_decomp) for loading, splitting, and batching multi-modal MRI scans.
• utils/pre_CT_MR.py, utils/pre_grey_rgb.py for modality normalization and format conversion.
b. Core Model & Training
• OurModel.py & networks/ (unet.py, fusion.py, region_to_location.py) define model architectures.
• segment_anything/ subpackage holds the SAM backbone (image_encoder, prompt_encoder, mask_decoder, transformer).
• loss_function.py encapsulates task-specific loss metrics.
• train_*.py scripts (train_Ours.py, train_CoarseSeg.py, train_multi_gpus.py, train_Ours_TTA.py) orchestrate training loops, using config_args.py for hyperparameters.
c. Inference & Evaluation
• MedSAM_Inference.py & MedSAM_Interface.py wrap SAM-based inference.
• gui.py provides a simple UI for interactive segmentation.
• comparisons/ holds side-by-side inference and training code for DeepLabV3+, nnU-Net, vanilla SAM—used for benchmarking.
• utils/demo.py and plot_figure.py for visualizing results, and test_sam_for_ours.py for sanity checks.
d. Extensions
• extensions/point_prompt, text_prompt, seg_3dnii_sparse_marker each add specialized prompting or 3D label interpolation.
e. Configuration & Packaging
• setup.py and medsam.egg-info/ for packaging.
• config_args.py centralizes CLI flags for dataset paths, model settings, training hyperparameters.
f. Workspace & Outputs
• work_dir/MedSAM stores trained checkpoints, logs, and output masks.
Relationships and data flow
• Data Modules feed preprocessed MRI tensors into the Training pipeline.
• Training pipeline instantiates OurModel or SAM+fusion, computes losses, and writes checkpoints to work_dir.
• GUI calls inference APIs for live segmentation.
• Comparison scripts operate in parallel to benchmark other architectures; outputs feed into utils/plot_figure.py.
Architectural patterns & design principles
• Modular layering: data/preprocessing → core model → training/inference scripts → evaluation.
• Separation of concerns: comparison, extension, and core logic live in distinct directories.
• Reuse of external foundation model (SAM) via subpackage rather than reimplementation.
• Config-driven CLI for reproducibility.
Technologies & frameworks
• Python 3.x, PyTorch (in segment_anything modeling files and networks/), Numpy (npz data lists), Click or argparse (config_args.py), Jupyter notebooks for tutorials, simple Tkinter or PyQt (gui.py).