FOCUS-3D

FOCUS-3D provides a user-friendly napari plugin for interactive 3D cell segmentation and analysis.

✨ What you can do

    🚀 Run automatic 3D cell segmentation with pretrained FOCUS-3D models
    ✏️ Manually correct segmentation errors
    🖱️ Perform one-click segmentation
    🧩 Prepare curated patches for human-in-the-loop fine-tuning
    🧠 Fine-tune FOCUS-3D models on your own data
    🔬 Reconstruct selected 3D cell instances
    📊 Compute quantitative statistics

📚 Examples & Tutorials

Visit our website for examples, tutorials, and usage instructions: 🌐 FOCUS-3D Website

☁️ Try FOCUS-3D Online

Want to try FOCUS-3D without installing anything? 🚀 FOCUS-3D Online on Hugging Face Spaces allows you to upload your own 3D fluorescence microscopy data for quick online inference and testing.

Installation

1. Create a new environment

conda create -n focus3d python=3.10 -y
conda activate focus3d

2. Install torch

For CUDA 12.x, replace cu12x with your specific CUDA-compatible PyTorch build. For example, for CUDA 12.6:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126

Please check the official PyTorch installation guide if you need another CUDA version.

3. Install FOCUS-3D

pip install -U "focus-3d[gui]"

4. Download pretrained model

Users can download the pretrained model for 3D segmentation: https://huggingface.co/Qinghua-thu/FOCUS-3D/.

5. Install detectron2 (optional, only for Linux model fine-tuning)

For Linux, FOCUS-3D requires Detectron2 for segmentation model fine-tuning. Please install Detectron2 according to the official guide:

https://detectron2.readthedocs.io/en/latest/tutorials/install.html

For most Linux environments, the official source installation command is:

python -m pip install --no-build-isolation 'git+https://github.com/facebookresearch/detectron2.git'

6. Launch napari

python -m napari

Recommended Workflow

Step 1 — Load and inspect the image

1. In napari, open Plugins -> 3D Segmentation (FOCUS-3D).
2. Load a raw 3D microscopy image:
   • use File -> Open Folder for a .zarr dataset, which is recommended for large volumes;
   • use File -> Open File(s) for .tif or .tiff images.
3. Open the Basic tab and use Display Settings when the raw image is difficult to inspect:
   • move the Minimum and Maximum sliders to adjust contrast;
   • click Auto for percentile-based contrast adjustment;

Step 2 — Run automatic 3D segmentation

1. Open the Segmentation tab.
2. In Run Segmentation, set the parameters that are most likely to vary between datasets:
   • Z Ratio — the physical Z-to-XY spacing ratio. Use 1.0 for isotropic data.
   • Output Path — the directory used for the segmentation result.
   • Checkpoint — the pretrained or fine-tuned checkpoint.
   • Cell radius (pixel) — the approximate cell radius in the XY plane.
   • Background intensity — patches or cells with grayscale values less than this value will be removed.
   • Min size (3D) and Max size (3D) — remove small or large instances.
3. Use Advanced only when you need to change the GPU, configuration file, normalization percentiles, patch stride, batch size, or stitching thresholds. See the complete menu reference for parameter definitions and defaults.
4. Click Run 3D Segmentation.

After inference, FOCUS-3D loads a label layer into napari. The inference outputs are saved to the specified output path in both TIFF and Zarr formats.

Step 3 — Inspect and curate the segmentation

1. Return to the Basic tab.

2. Click Enter Curation Mode, then click a cell label to select it.

3. Correct common errors:

   • use Add to Label to recover missing regions;
   • use Subtract from Label to remove incorrect regions;
   • use Add New Label for a missed cell;
   • use Delete Current Z for a slice-specific error;
   • use Delete All Z to remove an incorrect 3D instance;
   • use Delete Inside ROI (All Z) to remove multiple labels in a selected region.

   For detailed instructions on label-editing operations and keyboard shortcuts, see Manual Curation.

4. For labels stored in Zarr format, edits are written directly to the underlying Zarr data, so no separate save step is required. Labels loaded from TIFF are edited in memory and must be saved manually from the Save panel after curation.

Step 4 — Use one-click segmentation when needed

One-click segmentation can accelerate the curation.

1. Keep both the raw image and segmentation label layer loaded.
2. Open Segmentation -> One-click segmentation.
3. Click Enter Inactive Mode to load the local refinement model and activate interactive refinement.
4. Click the target cell in the viewer and inspect the updated label.
5. Click Exit Inactive Mode after finishing.

Step 5 — Analyze the segmentation results

Open the Analysis tab after the segmentation has been checked.

Reconstruct one selected cell

1. Select a non-background cell in the label layer.
2. Set the Z Ratio.
3. Click Reconstruct Selected Label.
4. Save the reconstructed mesh as .npz when needed.

Inspect the full volume in 3D

1. Set the physical Z Ratio.
2. Click Switch to 3D View.
3. Rotate and inspect the image and labels.
4. Click Switch to 2D View to return to slice navigation.

Run morphometry analysis

1. Set the physical voxel size in Z, Y, and X.
2. Choose an output folder.
3. Run one or more tasks:
   • Basic Information for cell morphology and optional intensity measurements;
   • Neighborhood Analysis for centroid-based local organization;
   • Contact Graph Analysis for face-touching cell relationships;
   • Clustering for feature-based cell grouping.
4. Use Show feature to map supported results back to the napari label volume.

Step 6 — Prepare training data and fine-tune the model

Fine-tuning consists of preparing annotated training data and then running the fine-tuning notebook.

A. Prepare training data

There are two ways to prepare the annotations:

• If the entire volume has been annotated: you do not need to use Calculate Valid Patches in napari. Provide the fully annotated raw image and label volume to the fine-tuning notebook. The notebook includes code to automatically crop the full volume into training patches and organize them into imagesTr and labelsTr.

• If annotating the entire volume is impractical: use the napari plugin to curate selected patches:

  1. Keep the raw image and corrected label volume loaded.
  2. Open Segmentation -> Finetune with Current Labels.
  3. Click Calculate Valid Patches.
  4. Choose a Patch ID and set the Save Path.
  5. Click Curate Selected Patch.
  6. Correct the labels in the patch viewer and click Save.

Curated patches are saved as:

<save_path>/
├── imagesTr/
│   ├── patch_0001.tif
│   └── ...
└── labelsTr/
    ├── patch_0001.tif
    └── ...

Use Clear Patch Boxes to remove the patch overlays when finished.

B. Run fine-tuning from the notebook

The fine-tuning notebook is not included in the pip installation and must be downloaded separately from the FOCUS-3D repository.

Open:

notebooks/02_finetune.ipynb

If using a fully annotated volume, first run the notebook section that crops it into training patches. If patches were already exported from napari, directly set the notebook to use the prepared imagesTr and labelsTr directories.

Run fine-tuning to obtain a new checkpoint, then return to Segmentation -> Run Segmentation, select the new checkpoint in the Checkpoint field, and run segmentation again.

The napari plugin can assist with patch preparation, while model fine-tuning itself is performed in the notebook.

Detailed Interface Reference

The complete descriptions of all controls are maintained in:

• Basic tab
• Segmentation tab
• Analysis tab
• Keyboard shortcuts
• Output files and folders

Issues

If you encounter a problem, please file an issue with a detailed description, relevant logs, and a minimal example when possible. You can also contact zhangqh24@mails.tsinghua.edu.cn. We also look forward to collaborating and jointly developing new algorithms, or customizing models based on your data.

Citing

@article{zhang2026focus,
  title={FOCUS-3D: Robust, generalizable volumetric cell segmentation for three-dimensional fluorescence microscopy},
  author={Zhang, Qinghua and Mu, Zeyu and Liu, Boqi and Chi, Yunfeng and Li, Donglin and Wang, Wenjuan and Ni, Jian-Quan and Wan, Yinan and Yu, Li and Navajas Acedo, Joaquin and Yu, Guoqiang},
  journal={bioRxiv},
  year={2026},
  doi={10.64898/2026.08.25.746907}
}