Skip to main content
. 2023 May 27;82(7):595–610. doi: 10.1093/jnen/nlad040

Table 2.

Overview of the most commonly used, open-source deep learning models designed for the task of microscopy image analysis

Project Reference Model architecture Use cases Advantages Open-source code
U-Net (76) U-Net Segmentation of various structures, very adaptive
  • Basis for most modern neural networks in image analysis

  • High customizability

Various
StarDist (11) Modified U-Net + polygonal representation of objects 2D and 3D object detection and segmentation, designed for complex shapes.
  • Polygonal representation of objects, high accuracy at complex shapes

  • Works well on crowded images

  • Fast and accurate segmentation

https://github.com/stardist/stardist
Cellpose (60) Modified U-Net + vector representation of objects 2D and 3D object detection and segmentation, continuously improved by the community training
  • Generalizable to various cell types

  • Size-independent cell segmentation

  • Robustness to image noise and variations in intensity

  • Growing community, continuously improved

https://github.com/MouseLand/cellpose
SplineDist (77) Modified U-Net (extended StarDist) Extension of the StarDist network, segmentation of objects with more complex shapes and curves
  • Spline-based representation of objects, allows for more complex shapes being accurately delineated

  • Captures objects with higher degree of variability in shape

  • Can handle overlapping and touching objects

https://github.com/uhlmanngroup/splinedist
EmbedSeg (78) ERFNet (convolutional network with residual connections) Instance segmentation of cells and nuclei in microscopy
  • Embedding-based segmentation approach

  • Robust to object occlusion and overlap

  • Small memory footprint enables researchers to run on less specific software (e.g. laptops)

https://github.com/juglab/EmbedSeg
CellSeg (79) R-CNN (region based convolutional network) Segmentation on highly multiplexed fluorescence images
  • Very detailed tutorials

  • Easy to use, out-of-the-box segmentation

  • Fast integration with other programming workflows due to low-code implementation

https://github.com/michaellee1/CellSeg

Every model can be used for the task it was originally trained for or retrained from scratch, given enough training data is present. All of the listed models are freely and publicly available, some are even provided with hands on tutorials to help new researchers get started using them. While this list is not exhaustive, the authors would recommend to start out with one of the models listed above, due to their ease of use and already proven accuracy at the task of micrograph analysis.