Install

Install with uv!

# install uv first
curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv --python 3.12
source .venv/bin/activate  
uv pip install -e .

Docker Image Building

Please put the checkpoint folder under /src/feature_extraction/work_dir

# Docker for LP and EAO works the same. Only feature extraction command, either extract_feat_LP.sh or extract_feat_EAO.sh differs
cd CT-NEXUS
docker build -f Dockerfile -t ctnexus .
docker save ctnexus:latest | gzip > ctnexus.tar.gz

Run Container

Place test_demo in the current directory.

docker load -i ctnexus.tar.gz
mkdir test_demo_outputs

## for Non-ROI disease
docker container run --gpus "device=0" -m 32G --name ctnexus --rm -v $PWD/test_demo/images/:/workspace/inputs/ -v $PWD/test_demo_outputs/:/workspace/outputs/ ctnexus:latest /bin/bash -c "sh extract_feat_LP.sh"

## for ROI disease
docker container run --gpus "device=0" -m 32G --name ctnexus --rm -e MASKS_DIR=/workspace/inputs/fg_masks/adrenal_hyperplasia -v $PWD/test_demo/:/workspace/inputs/ -v $PWD/test_demo_outputs/:/workspace/outputs/ ctnexus:latest /bin/bash -c "sh extract_feat_LP.sh"

Usage

Feature Extraction

Extract features from CT images using pretrained CT-NEXUS model:

Basic usage (Non-ROI disease):

python src/feature_extraction/extract_feat_LP.py \
  -i /path/to/input/images \
  -o /path/to/output/features \
  --checkpoint ./work_dir/CT-NEXUS/fold_all/checkpoint_final.pth

ROI-based disease (with masks):

python src/feature_extraction/extract_feat_LP.py \
  -i /path/to/input/images \
  -o /path/to/output/features \
  --masks_path /path/to/foreground/masks \
  --checkpoint ./work_dir/CT-NEXUS/fold_all/checkpoint_final.pth

Parameters:

• -i, --input: Input directory containing .nii.gz CT images (required)
• -o, --output: Output directory for extracted features (.h5 files) (required)
• --masks_path: Path to foreground masks for ROI-based diseases (optional)
• --checkpoint: Path to pretrained model checkpoint (default: ./work_dir/CT-NEXUS/fold_all/checkpoint_final.pth)
• --batch_size: Batch size for inference (default: 1)
• --dump_dir: Directory to save debug images and masks (optional)

Output:

• Features are saved as .h5 files in the output directory
• Each image produces one .h5 file containing the y_hat dataset with extracted features

Preprocessing pipeline:

• Images are normalized using Z-score normalization
• Resampled to isotropic 1mm spacing
• Non-ROI: Center cropped/padded to 320×320×320
• ROI: Center cropped to 160×160×160 based on mask, then padded to 160×160×160

Preprocessing

export nnssl_raw="/path/to/nnssl_raw"
export nnssl_preprocessed="/path/to/nnssl_preprocessed"
export nnssl_results="/path/to/nnssl_results"

nnssl_plan_and_preprocess -d 001 -c onemmiso -np 60

Pretraining CT-NEXUS:

• Stage 1:

nnUNet_n_proc_DA=16 nnssl_train 001 onemmiso -tr HuberMAETrainer_BS24 -p nnsslPlans -num_gpus 4

• Stage 2:

nnUNet_n_proc_DA=16 nnssl_train 001 onemmiso -tr AlignedHuberFTTrainer_MaxPool_BS20 -p nnsslPlans -num_gpus 4

Pretrained Weights and Docker

Pretrained weights and Docker image are available here