The basic way toautomatically identify cells is to use intensity-based segmentation techniquessuch as Otsu-based methods 2-4 or Watershed transformation 5. The initialOtsu method was applied to binarization of various images, but it was improvedby the binarization technique that is invariant to the brightness through localbinarization or adaptive binarization, so that the cell area of variousbrightness can be found from the microscope image. Such intensity-basedtechniques are easy to use, but it is difficult to expect good performance forcomplex background or splitting of adjacent cells with two or more overlapping.Recently, a method of using the distance map based on the Otsu method as aninitial value for Watershed has been proposed to improve the segmentationaccuracy 22.Energy-minimization based image segmentationtechniques can produce better results than the intensity based techniques inthe above-mentioned difficult environmental conditions. ACM (Active ContourModel) 67 is a representative energy-minimization technique usingarea-based partitioning 8 and edge-based partitioning 9, and shows goodperformance in noise image.
However, ACM-based techniques require initial pointsso that it is impossible to perform a full-automatic segmentation and the initialpoint affects the accuracy. Moreover, because of high computational complexitycaused by the iterative convergence process, it is not suitable for the purposeof finding multiple cells at the same time.Gradient flow tracking (GFT) 2324, graphcut10-14, and level set25-27 are energy minimization based techniques thatare more suitable than ACM-based techniques for multiple cell segmentation. GFTidentifies the cells by clustering the slope vector using the characteristicthat the slope vector around the cell becomes very congested, but in the noisyenvironment, the slope at the cell boundary is very small and the direction isunreliable. GC-based algorithms 10-1428-32 are widely used because they areguaranteed to find a global optimal solution for pixel boundaries betweendistinct regions. However, these methods may produce the boundaries of theuneven step-like shape or different from the boundary that perceived by human,so that a more advanced segmentation method are required.Recently,feature definition and classifier design have been suggested in various machinelearning-based cell segmentation techniques 15-19. Bayesian inference andKalman filter 17 showed good performance in cell segmentation and tracking.
Dynamicshape modeling (DSM) extended the state vector of a classical Kalman filter tocompute the morphological changes. The slope of the sigmoid function at thecell boundary is used as a probability model for shape inference by modelingthe uncertainty of the shape. The cell detection technique 18 using theclassifier trained from the seed provided as the ground truth is possible toclassify by binary support vector machine (SVM) in various datasets includinghistopathological images of breast cancer, fluorescence images of HEK (humanembryonic kidney) cells, and phase-contrast images of HeLa cells. Seeds shouldbe provided for each cell for learning, and a 92-dimensional histogram of thebrightness, shape and the brightness difference of the boundary is used as acharacteristic.
Supervised learning-based cell segmentation 15 can identifycells from histopathologic images using color and texture features extracted bylocal Fourier transforms. MDC (most discriminant color) space using intra-classand inter-class covariance matrices of local Fourier transforms for cell andnon-cell regions is better than RGB, and accuracy was improved by cellseparation through concave dot detection. However, this way of simply cuttingthe border of adjacent cells straight is different from the actual boundary ofthe cell. ilastik 16 classify the features of color, edge, and texture aroundeach pixel with a random forest classifier to detect the objects or regions. Theselearning-based image segmentation techniques generally have good segmentationability, but can not perform segmentation at a very precise level.The study ofdividing the connected cell region of interest (ROI) by assuming the ROI as a2D Gaussian mixture model (GMM) and using the expectation-maximization method producesresults close to the perceived boundaries of humans2021. However, since there are a plurality of parameters, it is necessaryto set the image data set depending on the image data set. Furthermore, thesegmentation quality is not good for a cell region in which a difference incontrast between the foreground and background regions is small.