Advances in automated image analysis
György Vereb · Cytometry Part A · 2014
Albert Szent-Györgyi, the Hungarian Nobel laureate who discovered vitamin C is widely quoted for having said: "Discovery consists of seeing what everybody has seen and thinking what nobody has thought." But do we always see enough to think what we need to think? This is becoming an increasingly emphatic enigma in today's research and diagnostics, with "high throughput" and "omics" evolving from an idea - through trendy - into the often optimal solution. In this special section, we have collected the latest papers on automated image processing that allow classification / extraction of biologically relevant data from large sets of images. These articles encompass three highly dynamic areas of automated image analysis: automatic segmentation and object detection, machine learning in target classification, and in depth analysis of subcellular morphological changes that can be used for high throughput pharmacological screening. Advanced digital pathology systems allow the digitization of an entire histological slide at very high resolution, affording in depth analysis of histologically diverse areas both from morphological and functional aspects 1. With such evolution of tissue cytometry, segmentation of cells and subcellular organelles has been much demanded, and consequently numerous developments in the field have followed. Even recently, several new approaches were proposed for segmenting nuclei 2, 3, as well as for the segmentation of moving cells partly based on segmenting their nuclei 4. One of the greatest challenges in segmentation procedures is the handling of touching or overlapping cells. In this issue, Arslan et al. (page 480) undertake to improve the existing algorithms of segmenting white blood cells in smears from peripheral blood and bone marrow, which is an important step in diagnostic evaluation, especially in leukaemias, where cellular aggregates are frequent in these smears. Their approach is to combine two transformations to define the most appropriate markers and marking function for the watershed algorithm generally used for these tasks. First they use an efficient color transformation best fitted to the chromatic characteristics of white blood cells in order to suppress background pixels while preserving cell pixel intensities. Then they use the image in the new color space to calculate the distance transformation representing shape characteristics of white blood cells. This double transformation approach allows for drastically reduced false negatives and false positives when compared to methods exploiting color clustering or conditional erosion. Another area where overlapping or aggregated cells cause problems is the counting of intracellular parasites. Neves et al. (this issue, page 491) propose a strategy for exploiting the concave regions of cellular contours where cells touch each other. The design intends to overcome performance issues experienced with watershed transformations, level-sets, or maximum-intensity linking. These are usually robust alternatives to cell segmentation, but exploit the fact that the intensity of stained cells usually monotonously changes from the core to the boundary, presenting a unimodal distribution, which is not the case with often multinucleate macrophages. The procedure starts out conventionally, using the blobs of nuclear staining to segment cytoplasmic areas around them with k-means clustering (which proved to be superior to multi-level thresholding). The excitement comes at this point, as after smoothing cellular contours, the concavities characteristic of the areas where two otherwise convex cells touch each other are detected and used to separate the cells, regardless of the number of nuclear blobs they contain. The approach is tested on fluorescence images of Leishmania-infected macrophages and proves to be superior to algorithms that inherently ignore the possibility of multiple nuclei. The code can be run in Matlab and is available at the Journal website. Circulating tumor cells are powerful indicators of metastatic tumors, their relapse, or inadequate response to therapy. Consequently, their identification and enumeration carry great diagnostic significance. However, false negatives or false positives can equally have disastrous effects on critical treatment decisions. In addition, the cells that need to be precisely identified are expected to be very rare. Given that a large proportion of such cells circulating in the body should be identified, the initial efforts to use flow or image cytometry on a limited volume of blood 5 were migrated to the construction of endovascular capture devices, that can, after having resided in the blood stream, be immunostained and imaged. In their pioneering work, Svensson et al. (this issue, page 501) implement machine learning using a naive Bayesian classifier (NBC) based on a probabilistic generative mixture model. Classification exploits the signal distribution in three fluorescence channels, establishing the fluorescence signature of circulating (and captured) tumor cells by the NBC. While performance is comparable to support vector machine (SVM) based classification, the new method also allows unsupervised learning and thus does not require labeled training data. With its fourth article, the special section jumps inside the cell, to provide a new look at the actin molecules that are organized into filaments in various distinct modes: lining the edges of cells, deposited at the protrusions of cells, forming internal stress fibers, and showing up as cytoplasmic punctate signals. Lockett et al. (this issue, page 512) apply an artificial neural network algorithm which is fed by image intensity and spatial anisotropy measurements to quantitatively classify these subcellular features constituted from F-actin. Carrying the system beyond the subjectivity of observation, this new quantitative pattern analysis allows the mathematical modeling of cytoskeletal changes and affords their statistical assessment. Given that many pathological conditions, as well as numerous drugs, directly or indirectly target cytoskeletal proteins, such objective testing has a vast potential both in understanding the correlation of cell fates and the dynamically changing cell morphology, as well as in the high throughput screening of drug candidates targeting the cytoskeleton.