Image processing for identification and quantification of filamentous bacteria in in situ acquired images
Among the different problems that can occur throughout activated sludge processes in wastewater treatment plants (WWTP) and thus hinder proper sludge settling, two of them are commonly caused by filamentous bacteria: bulking and foam formation [1]. These microorganisms are part of the activated sludge biocenosis and considered to create the backbone for sludge flocs, but problems can arise when they outcompete the floc-forming bacteria under specific conditions [2]. Typically, concentrations of filamentous bacteria have been monitored by means of human-performed, off-line and time demanding methods based on optical microscopy. Since these techniques are intrinsically subjective and susceptible to human errors, automated image analysis tools have been used for monitoring activated sludge processes [3].
As a disadvantage, these analysis methods still make use of discontinuous microscopy, a practice that requires sampling, transport and sample preparation before evaluation. Besides being labor-intensive, information may be altered during this process. There is also no consensus on the minimum sample size and number of images required for proper evaluation of activated sludge [10]. The in situ microscopy appears as a promising technique for an accurate, rapid and sampling-free monitoring of the abundance of filamentous bacteria under real environment conditions [11]. It can be directly installed in bioreactors or pipelines, eliminating sample collection and preparation of slides. Moreover, it allows an analysis with low statistical error, since large amounts of independent images (i.e. each image is equivalent to a new slide) can be acquired for each sample [12].
Current in situ microscopes are based on bright field microscopy. As a consequence, image processing techniques different from the ones used with phase-contrast illumination and fluorescence microscopy are required for the identification of filamentous bacteria. First, segmentation between flocs and filaments through direct brightness thresholding (e.g. Jenne et al. [5] and Amaral et al. [9]) is not viable for in situ acquired images.
Algorithms proposed for bright-field microscopy, as the ones described by Motta et al. [7], Mesquita et al. [8] and Lee et al. [13], are also based on some characteristics that are not true in the case of in situ acquired images. Most of them are designed for discontinuous microscopy, where proper dilution can be performed in cases of high concentration and background images can be acquired and subtracted for image enhancement. This is also the case of the in situ system proposed by Koivuranta et al. [14], where samples are diluted before evaluation. Moreover, these works have typically employed median filtering for differentiation between flocs and filaments, calculated using large window size (e.g., 25 (times) 25 [13]). In contrast, we describe an approach based on the Euclidean distance transform for a more refined thickness determination.
This paper introduces an image processing method for identification of filamentous bacteria from in situ acquired images. Unlike the reviewed studies, this method is suitable for real-time evaluation of images without any staining, phase-contrast or dilution technique. Moreover, as highlighted by Khan et al. [10], most algorithms for image analysis of activated sludge are described shortly and superfluously. They are rarely evaluated in terms of sensibility, accuracy and commonly present pre-defined parameter values that are not justified. In the present manuscript, the proposed method is explained in details, including an adaptation of the concept of ROC curves for evaluation and optimization of the developed algorithm. Finally, the algorithm here introduced provides an estimation of total extended filament length using geodesic distance transform for both pruning and extent measurement.