Automated Segmentation of Confocal Scanning Laser Microscopy Images - A Bioinformatic Approach - J. B. Xavier1, A. Schnell2, S. Wuertz2, R. Palmer3, D. C. White4, J. S. Almeida5 1ITQB, OEIRAS, PORTUGAL; 2Institute of. Water Quality Control and Waste Management, Garching, GERMANY; 3Natl Inst Dental Craniofacial Res - Natl Insts Health, Bethesda, MD; 4Center for Environmental Biotechnology, the University of Tennessee, Knoxville, TN; 5Medical University of South Carolina, Dept Biometry and Epidemiology, Charleston, SC

ABSTRACT - The segmentation of Confocal Laser Scanning Microscopy (CLSM) images consists of assigning image voxels (3D pixels) to the appropriate image elements (cell material, exopolysaccharides, distinct populations, liquid media, etc). The OTS (Objective Threshold Selection) algorithm was developed to cope with three-dimensional images, rendering possible the full automation of the measurement of biofilm volumetric features and morphological quantification. The new procedure was validated with both fixed and variable image acquisition settings, with images obtained from laboratory flowcell device systems. The resulting software application also includes image preprocessing through normalization for contrast and tilting. A web interface was developed to make this tool accessible to external users. The validation results and the interface for external access are presented.

Confocal laser scanning microscopy (CLSM) is the method of choice to capture structure of live biofilms due to its non-invasive and non-destructive character. Series of optical cross-sections, collected at different depths in order to scan a given volume of interest, provide extensive three-dimensional structural data. Time-course analysis of biofilm morphogenisis is also possible using CLSM, since it allows for non-destructive and repetitive visualization of live biological structures in their natural hydrated state.
The first step in most quantitative image analysis operations is image segmentation - the process of identification of different elements within the image. CLSM imaging in biofilm studies uses multiple color channels to record information on individual elements, such as cell material, EPS, distinct populations and liquid media. For each of the color channels, which is stored as a grayscale image, segmentation is commonly performed through direct thresholding at a given grayvalue – the threshold level. Manually setting a threshold level, which is usually performed interactively by an experienced operator, restrains the implementation of fully automated analysis procedures. The Objective Threshold Selection (OTS) is an automated method for the segmentation of CLSM three-dimensional data. It uses statistical analysis of the gradient of the 3D data provided by a CLSM image stack to set a threshold level, together with arbitrary confidence levels. This results in the automation of quantitative analysis procedures dependent on image segmentation, together with confidence levels on the direct measurements.

[click to see enlarged image]

A web-based application project is under development that will provide access to a whole range of automated image processing algorithms.This will include a CLSM biofilm image database, where registered users may submit images for analysis with the algorithms. Also, an updated database of registered users will provide a meeting point for researchers interested in exchanging information about CLSM biofilm imaging and quantitative analysis. Check out the project under development for demonstrations and updated news!

VALIDATION OF THE METHOD - Performance of the OTS procedure developed by the authors was evaluated by comparing the results obtained using this automated procedure with manual image segmentation by an experienced operator. The quantitative parameters determined include biovolume and interface area. The biovolume estimates by the two methods for a single species phenathrene degrading Sphingomonas sp. A0 biofilm are presented bellow. Confidence intervals for the OTS estimates correspond to a 95% probability level identified by propagation of threshold values.

For time resolved studies two microscope calibration methods were compared. In such studies, biofilm growth is monitored throughout time therefore image acquisition starts at low biomass quantities, which increase throughout time. For the first case, microscope contrast settings were selected at the moment of the acquisition instance (when cellular density is the lowest) and maintained throughout the entire experiment. As a result, data collected at later stages shows significant brightness saturation. For the second case the microscope settings were adjusted prior to each image acquisition, and therefore were tuned for the particular sample fluorescence.
The OTS method was found to be equally applicable to images obtained using both methods. The confidence levels behaved as expected, with the error associated with the segmentation showing a dependency of the cellular density in the first case and no evident dependency of the cellular density for the second case.

BIOFILM GROWTH MONITORING USING THE CSLM – Biofilms were obtained by inoculating laboratory flow cells. Staining methods such as commercially available vital florescent dyes or green fluorescence protein (GFP) expression were used for fluorescence image acquisition using CSLM. The choice of staining method is system specific and done according to the individual characteristics and the purpose of the analysis. Thin optical cross sections of the biofilm are imaged using CSLM. The entire three-dimensional biofilm structure was recorded by scanning along the biofilm depth, and the stacks of cross sections are stored as digital images.

THE OTS ALGORITHM – Objective Threshold Selection (OTS) is an extension of the robust automated threshold selection (RATS) algorithm. It uses statistical analysis of the brightness gradient of each color channel to determine a threshold level (T-level) for each separate color channel. The T-level is a weighed average of the brightness values of the voxels which attributes greater weight to high gradient voxels, i.e. where brightness shows steep changes.

where

The OTS algorithm considers the particular geometry of flowcell biofilm systems which is roughly planar. A statistical distribution of vertically oriented T-levels results in a cumulative distribution function F(T). A T-level can then be determined with arbitrary confidence levels, for example 95%.

Confidence levels can then be used for determining the error associated with the OTS for any quantitative analysis posterior to image segmentation. In this example, the confidence interval for T-level is determined from F(T), in the top plot, and propagated to find maximum and minimum values for the confidence interval of the volumetric parameter being estimated, in the bottom plot.

ACKNOLODGMENTS - This work is financially supported by the Foundation for Science and Technology/ M.C.T., PORTUGAL, through the grant PRAXISXXI/BD/18285/98 .

jxavier@itqb.unl.pt

Presented at the ASM 101st general meeting - Orlando, May 2001