nnU-Net Models for Paper "Bone Metastasis Detection on CT with Deep Learning Models Trained Using Multicenter, Multimodal Reference Standards: Development and Evaluation"

Model Description

This repository contains two nnU-Net models for automated bone metastasis detection in body CT scans, developed for the paper "Deep learning achieves expert-level bone metastasis detection on CT using multimodal reference standards" published in Radiology: Artificial Intelligence.

• Task 503 (Model 2): Trained on both CT-visible and CT-indeterminate bone metastases
• Task 504 (Model 1): Trained on CT-visible bone metastases only

Both models use the 3D low-resolution nnU-Net configuration and were trained with 5-fold cross-validation on contrast-enhanced abdominal and thoracic CT scans. Model 2 generally achieves better performance and is recommended for clinical applications. 

Installation

Install nnU-Net v1:

pip install nnunet

Usage

Setup Environment Variables

# Set the path to the downloaded model folder
export RESULTS_FOLDER="/path/to/downloaded/nnUNet_trained_models"

Run Inference

# Set input and output directories
INPUT_DIR="/path/to/your/ct/scans"
OUTPUT_DIR="/path/to/output/folder"

# Run prediction with Model 2 (recommended)
nnUNet_predict -i ${INPUT_DIR} \
               -o ${OUTPUT_DIR} \
               -t 503 \
               -m 3d_lowres \
               -f 0 1 2 3 4 \
               -tr nnUNetTrainerV2_DP \
               -p nnUNetPlansv2.1

# For Model 1, change -t 503 to -t 504

Input Requirements

• CT scans in NIfTI format (.nii.gz)

Output

The model generates segmentation masks in NIfTI format, where each detected bone metastasis is labeled as a separate region. 

Citation

If you use these models, please cite our paper: "Bone Metastasis Detection on CT with Deep Learning Models Trained Using Multicenter, Multimodal Reference Standards: Development and Evaluation" in Radiology: Artificial Intelligence.