We developed a tool dl4mic, that allows are users to train and run deep learning solutions from a graphical user interface integrated into FIJI/ImageJ. Currently 4 network-models are available within the tool: UNet, Dense-UNet, Noise2Void and Stardist. The tool allows to quickly integrate further models. For cell-segmentation we propose Cellpose. Cellpose is already trained with a large number of cell images and can often be used without further training. Our version of cellpose allows to run batch processing from within the graphical user interface.


Cellpose allows to segment cells of different types from different imaging modalities. The deep-learning network is pre-trained with images of 70000 cells. Cellpose uses a vector flow representation of cells. The input is  an image of the cells and optionally an image of stained nuclei.

MRI added a batch processing button, which allows to run the batch segmentation of cells from the graphical user interface instead of using the python from the command-line.

See also:

  • Cellpose (Presentation for the MRI image analysis group,  05.02.2020)
  • Cellpose reloaded (Presentation for the MRI image analysis group, 04.03.2021)


n2v zoom

Noise2Void is a deep-learning method that can be used to denoise microscopy images. No specific training datasets are required, only your noisy images. One noisy image is sufficient to train a network. Use dl4mic to train and apply the Noise2Void.


Automatic detection and segmentation of cells and nuclei using star-convex polygons.

Use dl4mic to train and apply StarDist.



UNet is a deep-learning method for pixel classification. It can be trained to segment different kinds of objects in images. The first half of the U-Net architecture is a downsampling convolutional neural network which acts as a feature extractor from input images. The other half upsamples these results and restores an image by combining results from downsampling with the upsampled images. Use dl4mic to train and apply the UNet.