ConfocalGN : a minimalistic confocal image simulator

by   Serge Dmitrieff, et al.

SUMMARY : We developed a user-friendly software to generate synthetic confocal microscopy images from a ground truth specified as a 3D bitmap with pixels of arbitrary size. The software can analyze a real confocal stack to derivate noise parameters and will use them directly to generate new images with similar noise characteristics. Such synthetic images can then be used to assert the quality and robustness of an image analysis pipeline, as well as be used to train machine-learning image analysis procedures. We illustrate the approach with closed curves corresponding to the microtubule ring present in blood platelet. AVAILABILITY AND IMPLEMENTATION: ConfocalGN is written in Matlab but does not require any toolbox. The source code is distributed under the GPL 3.0 licence on


Synthesizing Filamentary Structured Images with GANs

This paper aims at synthesizing filamentary structured images such as re...

BrainPainter: A software for the visualisation of brain structures, biomarkers and associated pathological processes

We present BrainPainter, a software that automatically generates images ...

New Methods to Improve Large-Scale Microscopy Image Analysis with Prior Knowledge and Uncertainty

Multidimensional imaging techniques provide powerful ways to examine var...

An effective and friendly tool for seed image analysis

Image analysis is an essential field for several topics of life sciences...

Generating Semi-Synthetic Validation Benchmarks for Embryomics

Systematic validation is an essential part of algorithm development. The...

Image Analysis in Astronomy for very large vision machine

It is developed a very complex system (hardware/software) to detect lumi...

FastPathology: An open-source platform for deep learning-based research and decision support in digital pathology

Deep convolutional neural networks (CNNs) are the current state-of-the-a...


Confocal microscopy allows the bioscience community to image biological processes in real time with a good spatial resolution, in particular with a higher contrast in the direction of the optical axis compared to other methods. However, the resolution on object shape is still limited by the diffraction limit, the shot noise, and possibly background fluorescence, tagging heterogeneity and fluorophore stochasticity tsien1995fluorophores ; pawley2006fundamental . Moreover, the resolution is anisotropic, typically several time worse in the optical axis than in the other axis. In many cases, these limitations can drastically hinder the ability to reconstruct a shape from an image.

To extract information from microscopy, image analysis tools are widely used. They usually involve image processing, such as background substraction and removal hwang1995adaptive , and deconvolution de2001image . Additional steps then can include segmentation and/or object recontructions using a wide variety of techniques. Methods to improve object reconstruction include model fitting, feature classification tomer2010profiling ; sommer2011ilastik , etc.

By altering the image and extracting features, image analysis can introduce artifacts pham2000current , that could remain undetected by the user when the analysis is fully automated. To assess the quality of the information obtained from imaging, it is thus necessary to know the quality of the image analysis procedure. This is however difficult because, especially in biological sample, the ground truth is not known, and the accuracy of the analysis cannot be quantified. This is why a need exist for a reliable software to generate simulated confocal microscopy data from a known ground truth.

Figure 1: A : illustration from confocal microscopy of a platelet microtubule marginal band severin2013confocal tagged with SIR-tubulin (top, data courtesy of A. Mathur) and a simulated confocal image of a ring (bottom). B : the ground truth, for image A is titled ring (black). The voxels above the Otsu threshold are shown in blue. C : histogram of voxel values for noise (red, n=23000 pixels) and signal (blue, n=1000 pixels), for the confocal microscopy image (dots) and for a simulated microscopy image (lines).

To simulate a confocal image, software needs to address both optical limitations and noise, i.e. convolve the ground truth with the point spread function of the microscope, and add noise - either before or after convolution according to the nature of the noise. Several software (the most notable being Huygens software) allow the simulation of point spread function (PSF), based on known the microscope components, i.e. the optical path. It is possible to measure the experimental PSF from microscopy images using similar software. Possible limitations include the non-uniformity of the refraction index in real samples and the aberrations that could stem from lens misalignment and imperfections.

Starting from a known PSF, a few software suites exist to simulate confocal imaging. They include the non-free Huygens software professional and the open-source microlith. While the later does not include a noise simulator, the former allows the user to add a background as well as a Gaussian or Poisson noise ; however, knowing the correct noise intensity and distribution is not straightforward. Each source of noise background noise, autofluorescence, shot noise) can also be theoretically simulated sheppard2006signal , but the resulting noise is a complex convolution of the different the noises and distortions sources (including the PSF), and can only simulated accurately only by complex models, and in very simplified experimental setup herberich2012signal .

Overall, any confocal simulator will be limited to a certain level of detail by practicality, and ultimately by the lack of information on the sample structure and properties - which is precisely the unknown. While such detailed simulators are extremely useful, simpler alternatives are possible, especially when it comes to testing image analysis software.


We developed ConfocalGN, a minimal solution to simulate confocal data from a ground truth, a measure of the PSF, and the noise and signal distribution.

The ground truth

is provided as a high resolution image, either as a MATLAB matrix, or as a .tiff file. A pixel value corresponds to no fluorescence, and pixel values are rescaled to match the desired mean signal value.


is entered as a three components vector containing the deviation of the Gaussian that fits best the PSF, since confocal PSF can efficiently be modeled as Gaussian. We do not provide a mean to compute the PSF since several third party software offer this possibility.

The noise distribution

is entered as the three first moments of the noise distribution measured experimentally. It is then modeled by a gamma distribution, as supported by experimental an theoretical evidence


, but a gaussian approximation is used if the user-provided noise has no skew (i.e. the third moment is zero).

The signal distribution

is entered as the mean value of the signal observed experimentally.

Interestingly, ConfocalGN offers the possibility to directly input a sample image, from which the noise and signal distribution will be calculated. This is extremely powerful as it allows for a realistic noise without an overwhelmingly complex simulation of the noise and distorsion sources. Additionally, the user can choose to use other software to convolve the ground truth image with the PSF, and use ConfocalGN only for noise modeling.

ConfocalGN offers a minimal segmentation function doing gaussian blurring and Otsu thresholding to discriminate noise vs signal pixels, but the user can provide his own segmentation function instead. In Fig. 1 A, we show a sample image (top) and an simulated image (bottom). Fig. 1 C, shows the histogram of values for the voxels under the Otsu threshold (noise, in red) and voxels above the threshold counting (signal, in blue).

ConfocalGN offers the user a very simple tool to test his image analysis software. The minimal input by the user is only (1) a ground truth image, (2) the confocal PSF, (3) a sample image (or the moments of the noise and signal distributions). Because of the approximation of the PSF and its simplified treatment of fluorophore, background and sensor noise, it is not meant to be an exact simulation. However, the complexity of biological systems, including intra-sample heterogeneity, inter-sample diversity and experimental variations mean that exact theoretical solutions are usually not achievable. Here, we rather chose to start from measured PSF and noise values to best address experimental conditions.


  • (1) R. Y. Tsien and A. Waggoner, “Fluorophores for confocal microscopy,” in Handbook of biological confocal microscopy, pp. 267–279, Springer, 1995.
  • (2) J. B. Pawley, “Fundamental limits in confocal microscopy,” in Handbook of biological confocal microscopy, pp. 20–42, Springer, 2006.
  • (3) H. Hwang and R. A. Haddad, “Adaptive median filters: new algorithms and results,” IEEE Transactions on image processing, vol. 4, no. 4, pp. 499–502, 1995.
  • (4) J. B. De Monvel, S. Le Calvez, and M. Ulfendahl, “Image restoration for confocal microscopy: improving the limits of deconvolution, with application to the visualization of the mammalian hearing organ,” Biophysical Journal, vol. 80, no. 5, pp. 2455–2470, 2001.
  • (5) R. Tomer, A. S. Denes, K. Tessmar-Raible, and D. Arendt, “Profiling by image registration reveals common origin of annelid mushroom bodies and vertebrate pallium,” Cell, vol. 142, no. 5, pp. 800–809, 2010.
  • (6) C. Sommer, C. Straehle, U. Köthe, and F. A. Hamprecht, “Ilastik: Interactive learning and segmentation toolkit,” in 2011 IEEE international symposium on biomedical imaging: From nano to macro, pp. 230–233, IEEE, 2011.
  • (7) T. Ideker, J. Dutkowski, and L. Hood, “Boosting signal-to-noise in complex biology: prior knowledge is power,” Cell, vol. 144, no. 6, pp. 860–863, 2011.
  • (8) T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham, “Active shape models-their training and application,” Computer vision and image understanding, vol. 61, no. 1, pp. 38–59, 1995.
  • (9) D. L. Pham, C. Xu, and J. L. Prince, “Current methods in medical image segmentation 1,” Annual review of biomedical engineering, vol. 2, no. 1, pp. 315–337, 2000.
  • (10) S. Severin, F. GAITS-IACOVONI, S. Allart, M.-P. GRATACAP, and B. Payrastre, “A confocal-based morphometric analysis shows a functional crosstalk between the actin filament system and microtubules in thrombin-stimulated platelets,” Journal of Thrombosis and Haemostasis, vol. 11, no. 1, pp. 183–186, 2013.
  • (11) C. J. Sheppard, X. Gan, M. Gu, and M. Roy, “Signal-to-noise ratio in confocal microscopes,” in Handbook of biological confocal microscopy, pp. 442–452, Springer, 2006.
  • (12) G. Herberich, R. Windoffer, R. E. Leube, and T. Aach, “Signal and noise modeling in confocal laser scanning fluorescence microscopy,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 381–388, Springer, 2012.