SAP (Skeletal Anatomical Pretraining)

This repo contains a clean, minimal reference for:

1. Running inference with the SAP skeletal model (formerly CATL) on whole-body images.
2. Fine-tuning SAP for downstream tasks using the IB-sampling patch-based loader.

The accepted paper PDF is in 1-s2.0-S2212137426000072-main.pdf.

Quick Start

1. Environment

Install core deps (PyTorch/MONAI should match your CUDA setup):

pip install -r requirements.txt

2. SAP Inference

Expected input layout (one folder per case):

input_dir/
  CASE_001/
    T1.nii.gz
    b1000.nii.gz
  CASE_002/
    T1.nii.gz
    b1000.nii.gz

Inference is driven by a JSON config. Edit sap/infer_config.json and run:

python sap/infer.py sap/infer_config.json

If no path is provided, sap/infer.py defaults to sap/infer_config.json.

Set model_type in the config to control the output classes:

1. \"skeletal\" → out_channels=10
2. \"lesion\" → out_channels=2 (background vs lesion)

Optional: save a probability map for a specific class by setting "save_probs": true and "prob_class": <idx> in the config.

python sap/infer.py \
  --input_dir /path/to/input_dir \
  --output_dir /path/to/output_dir \
  --weights /path/to/sap_checkpoint.pt \
  --save_probs --prob_class 1

3. SAP Fine-Tuning

The fine-tuning script uses the IB-sampling patch layout. A helper script to generate patches from whole-body volumes is included at utils/ib_sampling/prepare_dataset.py.

train_dir/
  lesion_patches/
    PATIENTID_T1_positive_0001.nii.gz
    PATIENTID_b1000_positive_0001.nii.gz
    PATIENTID_label_positive_0001.nii.gz
  background_patches/
    PATIENTID_T1_negative_0001.nii.gz
    PATIENTID_b1000_negative_0001.nii.gz
    PATIENTID_label_negative_0001.nii.gz

Fine-tuning is driven by a JSON config. Edit sap/finetune_config.json and run:

python sap/finetune.py sap/finetune_config.json

If no path is provided, sap/finetune.py defaults to sap/finetune_config.json.

Other initialization modes:

# Scratch
# Set \"training_mode\": \"Scratch\" in the config.

# SSL (self-supervised backbone)
# Set \"training_mode\": \"SSL\" and provide \"ssl_checkpoint\" in the config.

Dataloader (IB-sampling)

The patch-based loader is vendored from the IB-sampling repository and lives in utils/ib_sampling/. It comes from:

• https://github.com/jwutsetro/IB-sampling

The accompanying paper is:

• Instance-Balanced Patch Sampling for Whole-Body Lesion Segmentation (MICCAI 2025) https://doi.org/10.1007/978-3-032-08009-7_18

Notes

• SAP replaces the older name CATL.
• The scripts are kept minimal on purpose and avoid hard-coded paths.
• If your checkpoints come from torch.nn.DataParallel, the scripts automatically strip the module. prefix.