IMAGE SEGMENTATION METHODS FOR THE QUANTIFICATION OF CANDIDA CELLS IN FLUORESCENCE IMAGES

Acknowledgement References This work was supported by the Leibniz Association. Dietrich S.1,4, Brandes S.1,4, Hünniger K.2 , Engert N.3,4, Jacobsen I.3,4, Kurzai O.2,4, Figge T.1,4 Applied Systems Biology1, Fungal Septomics2, Microbial Immunology3 Leibniz Institute for Natural Product Reserach and Infection Biology - Hans Knöll Institute, Jena Friedrich Schiller University, Jena4 Introduction Candida species are commensals in the human body and do no damage under normal circumstances. However, in immunocompromised patients they can lead to severe infections. Our body posseses several defence mechanisms against these infections. For example polymorphonuclear neutrophils (PMNs) that are recruited to the site of infection and phagocytose and kill Candida cells or our natural skin barrier that builds the first line of defence against any infection. To study the interplay of PMNs or epithelial cells and fungal cells, phagocytosis assays or adhesion assays can be used. Fungal cells are fluorescently labeled and imaged. Manual analysis of these images can be time consuming. Therefore, we use automated image segmentation methods for the quantification of Candida cells. IMAGE SEGMENTATION METHODS FOR THE QUANTIFICATION OF CANDIDA CELLS IN FLUORESCENCE IMAGES Image data Segmentation evaluation [1] M. G. Netea et al., Immune defence against Candida fungal infections, Nat.Rev.Immunol. 15, 630-642 (2015) [2] N. Otsu, A threshold selection method from gray-level histograms, Automatica C (1), 62-66 (1975) [3] M. Farhan et al., A novel method for splitting clumps of convex objects incorporating image intensity and using rectangular window-based concavity point-pair search, Pattern Recognit. 46 (3), 741-751 (2015) [4] S. Brandes et al., Migration and Interaction Tracking for Quantitative Analysis of Phagocyte-Pathogen Confrontation Assays (submitted) Overview of immune defence mechanisms against Candida. Adapted from [1] Images from phagocytosis assays with PMNs in gray and Candida glabrata cells labeled in green. Basic image segmentation Advanced segmentation and classification Input: green fluorescence channel images converted to grayscale Output: segmented single cells and cell clusters - cluster splitting works good on small clumps, worse on large/occluded clumps - evaluation using manual segmentation of 30 images with 2045 cells: Adhesion assays with Candida albicans cells labeled in green on epithelial cells C. albicans Contrast enhancement Otsu's thresholding [2] Filtering noise (Gamma filter, γ = 0.3) (3x3 Median filter, 5x5 Morphological opening) Cluster splitting using concavity points [3]: 1) find concavity points in cell cluster 2) find corresponding point pairs for formation of cut lines 3) postprocess cut lines C. glabrata PMN Quantification Grayscale conversion No contrast enhancement necessary Adaptive thresholding Noise filtering through particle size (> 40 px) Cell classification using cluster skeleton 1) compute skeleton and distance transform 2) train SVM on labeled data with 4 classes (noise, single cells, cells with hyphae (chp), cell clusters) 3) classify test data +cell cluster single cell cell cluster Classification result # cells (man. Seg.) # cells (aut. Seg.) TP FP FN Recall Precision 2045 1927 1894 78 0.93 0.96 151 The image segmentation allows the automated measurement of fungal growth over time. In combination with segmentation and tracking of PMNs, phagocytosis rates can be measured [4]. In combination with segmentation of epithelial cells, the number of adherent fungal cells can be counted. The classification of cells with hyphae is the first step in measurement of hyphal growth and quantification of invasion behaviour under different environmental 0 10 20 30 40 50 60 Fungal growth (%) Time (min) q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q Goodness of fit: 1e+00 Weighted 5−P logistic regr. (nplr package, version: 0.1.4) 0.0 0.1 0.3 0.4 0.5 0.6 Relative fungal growth over one hour. Taken from [4] - classification can distinguish well between noise and single cells - classification of hyphae and clusters needs to be improved - classification on 176 objects using cross validation (k=4, 10 runs) - chp = cells with hyphae True class Identified class noise cell chp cluster noise cell chp cluster 0.908 0.068 0 0.022 0 0.976 0 0.024 0 0.1 0 0.9 0.053 0.111 0 0.836 % Contact: [email protected] Candida species are ubiquitous and can lead to severe infections. Our body possesses several defence mechanisms against infections, like cells of the innate immune syste m. To study the interplay of body cells and pathogens, different biological assays can be used, wherein fungal cells are labeled and imaged. We use automated image segmentation methods and machine learning to quantify Candida cells, thereby characterizing the cell interactions. Abstract