[CVPR2025] LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging

This repository contains the official implementation of our CVPR 2025 paper:

🎯 LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging

The paper introduces a novel framework for zero-shot lesion segmentation and longitudinal tumor tracking in 3D full-body imaging. By combining a large-scale lesion dataset, promptable segmentation, and deep-learning-based image registration, our framework achieves state-of-the-art results for both tasks.

Authors: Maximilian Rokuss, Yannick Kirchhoff, Seval Akbal, Balint Kovacs, Saikat Roy, Constantin Ulrich, Tassilo Wald, Lukas T. Rotkopf, Heinz-Peter Schlemmer, and Klaus Maier-Hein
Preprint:

News/Updates:

• 💻 4/25: LesionLocator code released! 🥳 The checkpoint can be found here
• 🗃️ 3/25: Lesion Dataset with Synthetic Follow-Ups Released 👉 here
• 📄 2/25: CVPR Acceptance 🎉

---

🛠️ Installation

1. Clone the repository

git clone https://github.com/MIC-DKFZ/LesionLocator
cd LesionLocator

2. Create a conda environment (recommended)

conda create -n lesionlocator python=3.12 -y
conda activate lesionlocator

3. Install the package

pip install -e .

If you want to use the napari viewer for visual inspection:

pip install "napari[all]"

---

🚀 Features & Usage

LesionLocator has two main modes:

1️⃣ Zero-Shot Lesion Segmentation (Single Timepoint)

Perform universal lesion segmentation using point or 3D bounding box prompts. No lesion type–specific fine-tuning required.

Usage

LesionLocator_segment
    -i /path/to/image(s)
    -p /path/to/prompt(s) 
    -t box 
    -o /path/to/output/folder
    -m /path/to/LesionLocatorCheckpoint 

Arguments

Argument	Description
-i	Input image(s): Path to an image file (.nii.gz) or a folder containing multiple images.
-p	Prompt(s): Can be either .json files containing 3D point or box coordinates, or .nii.gz instance segmentation maps. Prompts must have the same filename as their corresponding input image. Binary masks will be automatically converted to instance maps. More info here.
-t	Prompt type: Choose between point or box. Determines how the model interprets the prompts. Default is box.
-o	Output folder: Path where the predicted segmentation masks will be saved. Created automatically if it doesn't exist.
-m	Model folder: Path to the downloaded LesionLocatorCheckpoint directory containing trained model weights.
-f	Model folds: Specify one or more folds. Defaults to all 5 folds for ensemble prediction.
--disable_tta	Disables test-time augmentation (TTA) using mirroring. Speeds up inference at the cost of accuracy.
--continue_prediction, --c	Continues a previously interrupted run by skipping existing output files.
--visualize	Opens results in a napari viewer for inspection. Requires napari[all] installed.

🧠 If you provide a 3D (instance) segmentation mask as a prompt, LesionLocator will internally extract the bounding boxes or points automatically, depeding on the specified prompt type -t. Details on how to handle promting and the .json format can be found here. You can also run LesionLocator_segment -h for help.

Examples:

LesionLocator_segment -i image.nii.gz -p label.nii.gz -t box -o /output/folder -m /path/to/LesionLocatorCheckpoint 

LesionLocator_segment -i image.nii.gz -p points.json -t point -o /output/folder -m /path/to/LesionLocatorCheckpoint 

LesionLocator_segment -i /image/folder -p /prompts/folder -t box -o /output/folder -m /path/to/LesionLocatorCheckpoint 

---

2️⃣ Lesion Segmentation & Tracking Across Timepoints (Longitudinal)

Includes the lesion segmentation above but also tracks lesions across multiple timepoints using previous labels or prompts.

Usage

LesionLocator_track
    -bl /path/to/baseline.nii.gz
    -fu /path/to/followup1.nii.gz /path/to/followup2.nii.gz
    -p /path/to/baseline_prompt_or_mask(s)
    -t prev_mask
    -o /path/to/output
    -m /path/to/LesionLocatorCheckpoint

Arguments

Argument	Description
-bl	Baseline scan: Path to the .nii.gz image used as the baseline. Must match format expected by the model (or dataset config).
-fu	Follow-up scan(s): One or more .nii.gz files representing follow-up timepoints. Provide multiple paths separated by space for autoregressive tracking.
-p	Prompt: Path to the baseline prompt. Can be one .nii.gz instance segmentation mask or list of several semantic segmentation masks. Can also be a .json file with a point/box prompt. Used to initialize tracking. More info here.
-o	Output folder: Where tracked lesion segmentations will be saved. Created automatically if it doesn't exist.
-t	Prompt type: Options: point, box, or prev_mask. Use point/box to trigger segmentation of baseline lesions before tracking. Use prev_mask to track from an existing baseline label mask directly. Default is prev_mask.
-m	Model folder: Path to the LesionLocatorCheckpoint directory containing tracking model weights.
-device	Device selection: Options are cuda, cpu, or mps (Apple Silicon). Controls where the inference runs. Default is cuda.

You can also run LesionLocator_track -h for help.

Examples

This will do both the segmentation of the baseline scan using box prompts generated from baseline_label.nii.gz and then track each lesion:

LesionLocator_track -bl baseline.nii.gz -fu followup.nii.gz -p baseline_label.nii.gz -t box -o output/folder -m /path/to/LesionLocatorCheckpoint

Similarly works with .json prompts:

LesionLocator_track -bl baseline.nii.gz -fu followup.nii.gz -p points.json -t point -o output/folder -m /path/to/LesionLocatorCheckpoint

This will trigger just the tracking (-t prev_mask) of all provided semantic masks autoregressively for all follow-ups:

LesionLocator_track -bl baseline.nii.gz -fu followup1.nii.gz followup2.nii.gz -p mask1.nii.gz mask2.nii.gz mask3.nii.gz -t prev_mask -o output/folder -m /path/to/LesionLocatorCheckpoint

To run inference with LesionLocator, you need to download the pretrained model checkpoint:

🔗 Download LesionLocator Checkpoint

Once downloaded, extract the contents and use the -m argument in the CLI tools to point to the directory named LesionLocatorCheckpoint. Let us know if you run into issues downloading or loading checkpoints!

---

🗃️ Lesion Dataset with Synthetic Follow-Ups

We have released the Lesion Dataset with Synthetic Follow-Ups here (ca. 700 GB), which includes simulated follow-up scans with consistent instance labels. Due to image size, quality, or licensing constraints, not all images were used to synthesize a second timepoint and were excluded from this dataset. We reccomend downloading e.g. with wget since our servers can sometimes be a bit unstable.

wget -c --no-check-certificate -O LesionLocator_SynteticLongitudinalDataset.zip https://doi.dkfz.de/10.6097/DKFZ/IR/E230/20250324_1.zip

⚠ Important Notes

• Instance-based labels (not semantic): Lesions in a single patient scan are labeled consecutively, and the same lesion retains the same label across both timepoints.
• Synthetic deformations & segmentation masks: Some images contain unrealistic deformations or challenging segmentation masks, which may serve as useful test cases for improving automated analysis methods.
• Recommended use: This dataset is ideal for pretraining or for use alongside real longitudinal data to enhance model robustness and generalization.
• Longitudinal tracking tip: When using this dataset for longitudinal tracking, we highly recommend cropping one image during data augmentation to increase (translational) misalignment and better simulate real-world conditions.

Included Datasets

This dataset incorporates lesion data from various publicly available sources:

• COVID-19 CT Lung
• KiTS
• LIDC
• LNDb
• MSD Colon
• MSD Hepatic Vessels
• MSD Liver
• MSD Lung
• MSD Pancreas
• NIH Lymph
• NSCLC Radiomics

📜 This dataset is released under the CC BY-NC-SA 4.0 license.

---

More Related Code Releases

If you are interested check out nnInteractive or LongiSeg which we both just released:

• nnInteractive is our recent model for 3D interactive segmentation across all modalities, all target structures using various types of prompts! Check out the python backend or the napari plugin.    

• LongiSeg is an extension of the popular nnU-Net framework, designed specifically for longitudinal medical image segmentation. By incorporating temporal information across multiple timepoints, LongiSeg enhances segmentation accuracy and consistency, making it a robust tool for analyzing medical imaging over time. LongiSeg includes several methods for temporal feature merging, including the newly introduced Difference Weighting Block.    

---

📚 Citation

If you find our work useful, please consider citing:

@inproceedings{rokuss2025lesionlocator,
      title={LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging}, 
      author={Maximilian Rokuss and Yannick Kirchhoff and Seval Akbal and Balint Kovacs and Saikat Roy and Constantin Ulrich and Tassilo Wald and Lukas T. Rotkopf and Heinz-Peter Schlemmer and Klaus Maier-Hein},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
      year={2025},
      eprint={2502.20985}
}

---

📬 Contact

For questions, issues, or collaborations, please contact:

📧 maximilian.rokuss@dkfz-heidelberg.de