Selection and Tuning of a Fast and Simple Phase-Contrast Microscopy Image Segmentation Algorithm for Measuring Myoblast Growth Kinetics in an Automated Manner

Abstract

Acquiring and processing phase-contrast microscopy images in wide-field long-term live-cell imaging and high-throughput screening applications is still a challenge as the methodology and algorithms used must be fast, simple to use and tune, and as minimally intrusive as possible. In this paper, we developed a simple and fast algorithm to compute the cell-covered surface (degree of confluence) in phase-contrast microscopy images. This segmentation algorithm is based on a range filter of a specified size, a minimum range threshold, and a minimum object size threshold. These parameters were adjusted in order to maximize the F-measure function on a calibration set of 200 hand-segmented images, and its performance was compared with other algorithms proposed in the literature. A set of one million images from 37 myoblast cell cultures under different conditions were processed to obtain their cell-covered surface against time. The data were used to fit exponential and logistic models, and the analysis showed a linear relationship between the kinetic parameters and passage number and highlighted the effect of culture medium quality on cell growth kinetics. This algorithm could be used for real-time monitoring of cell cultures and for high-throughput screening experiments upon adequate tuning.

Monitoring of adherent live cells morphology using the undecimated wavelet transform multivariate image analysis (UWT‐MIA)

“…The evolution of the average cell counts and the average CCSR for both wells are shown in Figure A. A positive correlation is observed between the average CCSR curves and the average cell counts, both media supporting similar growth kinetics thus confirming the results presented in Juneau et al (). In Figure A, both curves (CCSR and cell count estimate) provide complementary information, as the CCSR is an estimate of the surface covered by the cells.…”

“…The method can be used on large datasets of PCM live‐cell images to assess cellular morphology and growth kinetics as a function of culture conditions and cell type without single‐cell segmentation. The CCSR can be used to assess growth kinetics of cells in PCM, providing kinetic parameters estimates that are similar to published data (from manual cell counts) for myoblasts (Juneau et al, ). The total cell count estimates obtained from UWT‐MIA could be used for performing similar growth kinetics studies, since these estimates are highly correlated with the CCSR.…”

“…The impact of the culture medium used in the current work was already assessed qualitatively: myoblasts grown in SFM (LOBSFM) contained more granules, formed more vacuoles and were more elongated (more delicate) than myoblasts grown in SSM, showing a certain impact of the culture media on cell state. However, it was shown quantitatively that the SFM was providing similar cellular growth kinetics than SSM (Juneau et al, ; Parent et al, ), while maintaining cellular phenotype. For both SFM and SSM, this qualitative assessment was confirmed in a quantitative manner by the proposed approach based on UWT‐MIA and PCM images.…”

Segmentation of phase contrast microscopy images based on multi-scale local Basic Image Features histograms

“…The trainable segmentation approach (using multi-scale local BIFs histograms) was also compared with specialised PCM image segmentation algorithms based on contrast filters (Topman et al 2011; Juneau et al 2013; Jaccard et al 2014). Trainable segmentation outperformed two of the three approaches and produced results approaching those obtained using the third (best performing) one (Table 1).…”

“…Otsu's) do not usually produce satisfactory results. Specialised segmentation approaches that rely on a priori knowledge of the structure and properties of PCM images have been developed, including methods based on contrast filters (Bradhurst et al 2008; Topman et al 2011; Juneau et al 2013; Jaccard et al 2014), active contours (Ambühl et al 2012; Seroussi et al 2012), weak watershed assemblies (Debeir et al 2008) and image formation models (Yin et al 2012). More recently, trainable segmentation methods for microscopy images based on statistical learning of image features (e.g.…”