Type of presentation: Oral

ID-9-O-2737 Quality control in cell image segmentation and tracking using simulated cell images

Kozubek M.¹, Svoboda D.¹, Ulman V.¹

¹Centre for Biomedical Image Analysis, Masaryk University, Brno, Czech Republic

kozubek@fi.muni.cz

Cell imaging using optical microscopy still faces the problem of the quality of cell image analysis results. The majority of acquired 2D as well as 3D cell image data are typically of not very good quality due to degradations caused by cell preparation, optics and electronics. That is why image processing algorithms applied to this data often yield imprecise and unreliable results. As the ground truth (GT) for given image data is typically not available the outputs of different image analysis methods can be neither verified nor compared to each other. This problem can be partially solved by estimating GT by experts in the field. However, in many cases such GT estimate is very subjective and strongly varies among different experts.

In order to overcome these difficulties we have created a toolbox that can generate 3D models (so-called digital phantoms) of artificial biological objects (primarily cell nuclei, see Fig. 1) along with their corresponding images virtually degraded by specific optics and electronics (using simulation of optical system and electronic detection, see Fig. 2). Image analysis methods can then be tried out on such synthetic image data. The analysis results (such as segmentation or tracking results) can be compared (e.g., using a difference) with GT derived from input models of a particular type of objects. In this way, image analysis methods can be compared to each other, their quality can be evaluated and their development can be fostered.

We started our development with simulations of cell nuclei of HL60 cells (to model simple isolated round objects) and cell nuclei of granulocytes (to model complex shapes) [1]. Then we paid attention to clusters of cells and their formations in tissues [2]. Finally, we started working on time-lapse simulations [3]. Recently, we have also generated special benchmark data sets of simulated 2D as well as 3D time-lapse series that were employed in Cell Tracking Challenges at International Symposium on Biomedical Imaging in 2013 and 2014 [4]. In this way, several state-of-the-art image segmentation and cell tracking methods could be compared [5]. The simulation toolbox is freely available on Internet and is accessible via simple web interface [6].

References:

[1] D Svoboda, M Kozubek and S Stejskal, Cytometry 75A (2009), p. 494-509.
[2] D Svoboda, O Homola and S Stejskal, Proceedings of 8th International Conference on Image Analysis and Recognition, LNCS 6754 (2011), p. 31-39.
[3] D Svoboda and V Ulman, Proceedings of 9th International Conference on Image Analysis and Recognition, LNCS 7325 (2012), p. 473-482.
[4] http://www.codesolorzano.com/celltrackingchallenge
[5] M. Maška et al., Bioinformatics (2014), published online, doi: 10.1093/bioinformatics/btu08
[6] http://cbia.fi.muni.cz/simulator

The authors gratefully acknowledge funding from the Grant Agency of the Czech Republic under grant number P302/12/G157.

Fig. 1: First phase of the creation of synthetic image data – digital phantom generation for a cell nucleus of granulocyte and a cell nucleus of HL60 cell.

Fig. 2: Second and third phase of the creation of synthetic image data – blurring using virtual microscopy and noise introduction using light detector simulation for digital phantoms from Figure 1.