Table 4 A summary of segmentation assumptions in the surveyed literature

Biological assumptions

Image Contrast

Strong staining to get high SNR for actin fibers

[113]

Optophysical principle of image formation is known

[44]

Cell brightness significantly higher than background

[114, 115]

Cell signal higher than noise level in an acquired z-stack

[49, 116–118]

Object Shape

Biological assumptions about mitotic events like mother roundness and brightness before mitosis

[119–122]

Nucleus shape is round

[123]

Specifically designed for dendritic cells

[83]

Cell line falls into one a few object models. Cell must have smooth borders. E.coli model assumes a straight or curved rod shape with a minimum volume darker than background. Human cells assume nearly convex shape.

[124]

Cells posses only one nucleus

[125]

Algorithmic assumptions

Background/Foreground Boundary

Initializing level sets functions based on k-means clustering

[126]

Background

Background intensities are between the low and high intensities in the image

[127]

Local background must be uniform

[128, 129]

Background is piecewise linear and its intensities are between the low and high intensities in the image

[130]

Foreground

Clear distinction between touching cell edge pixel intensities

[122]

Foreground pixels are drawn from a different statistical model than the background pixels

[131]

Features computed based on their gray-scale invariants

[132]

Time

The first image of a time sequence should be segmented first by another algorithm like watershed

[69]

Intensity Distributions

Image pixel intensities follow bi-modal histogram

[42]

The statistical properties of the foreground and background are distinct and relatively uniform & foreground is bright, while the background is dark

[133]

Foreground and background follow Gaussinan distribution

[134]

Image pre-processing

Background flatfield correction

Image pre-processing: such as correcting inhomogeneous illuminated background intensities using a machine learning based approach to resolve differences in illumination across different locations on the cell culture plate and over time

[81]

Filters

Smoothing the image using Gaussian filter

[132]

Downsampling (binning) the images

[64]

Image smoothing and automatic seed placement are used

[56]

Hessian-based filtering for better cell location and shape detection

[44]

Non-linear transformation

Image pre-conditioning where the image is transformed to bright field before applying the threshold

[48]

Manual input

Manual interactivity is needed to compute segmentation

[84]