StarDist
A Python library for 2D/3D object detection and instance segmentation in microscopy images using star-convex shapes.
What is StarDist?
StarDist is a deep learning library for object detection and instance segmentation in 2D and 3D microscopy images. It uses a star-convex shape representation to model objects like cells and nuclei, solving the problem of accurately delineating individual instances in dense or complex biological images. The method is based on predicting distances to object boundaries along fixed rays, followed by non-maximum suppression to produce final segmentations.
Bioimage analysts, computational biologists, and researchers working with microscopy data who need to segment and quantify cells, nuclei, or other objects in 2D or 3D images. It is also suitable for developers building custom image analysis pipelines in Python.
Developers choose StarDist for its robust performance on challenging microscopy datasets, its support for both 2D and 3D data, and its integration with popular bioimaging tools. Its star-convex shape prior provides a good trade-off between flexibility and regularity, often outperforming generic segmentation methods for biological objects.
Overview
StarDist - Object Detection with Star-convex Shapes
Use Cases
Best For
- Segmenting nuclei in fluorescence microscopy images
- Analyzing histopathology images with H&E staining
- 3D cell detection and segmentation in volumetric microscopy data
- Quantifying cell populations in dense tissue samples
- Multi-class instance segmentation (e.g., classifying different cell types)
- Benchmarking instance segmentation algorithms in bioimaging
Not Ideal For
- Real-time analysis of live-cell imaging videos, due to computational intensity and lack of optimization for speed.
- Segmentation of non-star-convex objects like neurons or blood vessels, which have elongated or branching shapes.
- Projects needing minimal software dependencies, as StarDist requires TensorFlow and can have complex installation on some platforms.
- General computer vision tasks outside biomedical imaging, such as natural image or video segmentation.
Pros & Cons
Pros
Models objects as star-convex polygons or polyhedra, providing a balance between flexibility and regularity that excels at segmenting irregular cells in dense environments, as demonstrated in the MICCAI and WACV papers.
Handles both 2D images and 3D volumetric data, making it suitable for a wide range of microscopy techniques, from fluorescence to histopathology, with dedicated workflows in the example notebooks.
Includes out-of-the-box models for common tasks like nuclei segmentation in fluorescence and H&E images, allowing quick application without training, as listed under 'Pretrained Models for 2D'.
Supports classification of detected instances into discrete classes, enabling cell type differentiation alongside segmentation, as shown in the ISBIC paper and multi-class example notebook.
Cons
Requires careful setup of TensorFlow with specific CUDA/cuDNN versions for GPU support, and compilation issues are common on macOS and Windows, as detailed in the troubleshooting section with platform-specific fixes.
The method assumes objects are star-convex, which may not hold for all biological structures like filamentous networks, potentially reducing accuracy without custom modifications.
Relies on TensorFlow and other libraries like gputools for optimal performance, making it resource-intensive and potentially overkill for simple segmentation tasks compared to lighter tools.
Frequently Asked Questions
Related Projects
Cellposea generalist algorithm for cellular segmentation with human-in-the-loop capabilities
HoVer-NetSimultaneous Nuclear Instance Segmentation and Classification in H&E Histology Images.
MicroSAMSegment Anything for Microscopy
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.