Skip to content

Latest commit

 

History

History
164 lines (103 loc) · 6.82 KB

README.md

File metadata and controls

164 lines (103 loc) · 6.82 KB

EPFL Center for Imaging logo

🫁 Lung tumor nodules segmentation in mice CT scans

We provide a neural network model for lung tumor nodule segmentation in mice. The model is based on the nnUNet framework which we used in the full resolution 3D configuration (3d_fullres).

The goal of our tool is to facilitate the annotation of individual lung tumor nodules in mouse CT scans. The U-net model produces a binary mask representing the foreground tumor class. The tumor nodules are individually labeled based on the connected components method.

This project is part of a collaboration between the EPFL Center for Imaging and the De Palma Lab.

Input data specifications

Make sure that your input data is compatible with our model. To check the integrity of your input data, read Input data specifications.

Hardware requirements

Installing PyTorch with CUDA support and using a GPU for inference is strongly recommended. We report the following runtimes for inference on GPU and CPU, respectively:

  • GPU (RTX 3060, 12 GB RAM): 12 sec.
  • CPU (AMD Ryzen 9 5900X (Zen 3, 64MB L3), 12 Threads): 68 sec.

Installation

We recommend performing the installation in a clean Python environment. If you are new to Python, read our beginner's guide to learn how to do that.

The code requires python>=3.9, as well as pytorch>=2.0. If wish to use a GPU with CUDA support, you may want to install Pytorch first and separately following the instructions for your platform on pytorch.org.

Install mousetumornet using pip after you've installed Pytorch:

pip install mousetumornet

or clone the repository and install with:

git clone git+https://gitlab.epfl.ch/center-for-imaging/mousetumornet.git
cd mousetumornet
pip install -e .

Models

The model weights (~461 MB) are automatically downloaded from Zenodo the first time you run inference. The model files are saved in the user home folder in the .nnunet directory.

New versions of the model, trained on more annotated data, faster, or more performant, are to be released in the future. As of November 2023, the available models are:

These models are all available for use in our package and can be selected by the user (see Usage).

Usage in Napari

Napari is a multi-dimensional image viewer for python. To use our model in Napari, start the viewer with

napari

To open an image, use File > Open files or drag-and-drop an image into the viewer window. If you want to open medical image formats such as NIFTI directly, consider installing the napari-medical-image-formats plugin.

Sample data: To test the model, you can run it on our provided sample image. In Napari, open the image from File > Open Sample > Mouse lung CT scan.

Next, in the menu bar select Plugins > mousetumornet > Tumor detection. Select a model and run it on your selected image by pressing the "Detect tumors" button.

To inspect the results, you can bring in a table representing the detected objects from Plugins > napari-label-focus > Data table. Clicking on the data table rows will focus the viewer on the selected object.

Usage as a library

You can run a model (v1 in the example below) in just a few lines of code to produce a segmentation mask from an image (represented as a numpy array).

from mousetumornet import predict, postprocess

binary_mask = predict(your_image, model='v1')
instances_mask = postprocess(binary_mask)

Usage as a CLI

Run inference on an image from the command-line. For example:

mtn_predict_image -i /path/to/folder/image_001.tif/ -m <model_name>

The <model_name> should be in the model list, for example v42.

The command will save a mask next to the image:

folder/
    ├── image_001.tif
    ├── image_001_mask.tif

Run inference in batch on all images in a folder:

mtn_predict_folder -i /path/to/folder/ -m <model_name>

Will produce:

folder/
    ├── image_001.tif
    ├── image_001_mask.tif
    ├── image_002.tif
    ├── image_002_mask.tif

Usage recommendation

For use in a scientific context, we believe the model outputs should be considered as an initial guess for the segmentation and not as a definitive result. Certain deviations in the the instrumentation, acquisition parameters, or morphology of the tumor nodules, among other things, can affect performance of the model. Therefore, the detections should always be reviewed by human experts and corrected when necessary.

Dataset

Our latest model was trained using a dataset of 402 images coming from 17 different experiments and validated on 101 images. Four experts from Prof. De Palma's lab in EPFL participated in the data annotation. The images were acquired over a period of about 18 months using two different CT scanners. In this dataset, we report a dice score performance of around 0.63 on the validation images.

Contributing

Contributions are very welcome. Please get in touch if you'd like to be involved in improving or extending the package.

Issues

If you encounter any problems, please file an issue along with a detailed description.

License

This project is licensed under the BSD-3 license.

This project depends on the nnUNet library which is licensed under Apache-2.0.

Carbon footprint of this project

As per the online tool Green algorithms, the footprint of training the mouse tumor net model was estimated to be 105 g CO2e.

Related projects