We developed a system that automates the detection of nuclei in a set of 4D DIC microscope images of C. elegans embryos. One major advantage of this system is the use of local image entropy to quantify the appearance of the nucleus in the images. Our previous system used edge detection operators to quantify the appearance of the nucleus . Because these operators were sensitive to differences in image quality (e.g., brightness, contrast) among sets of images, the previous system required laborious hand-tuning of system parameters each time a new image set was used (see Additional file 1). Local image entropy is not sensitive to differences in image quality among sets of images because it represents the smoothness of the image texture (see Additional file 1). Therefore, our system can be applied to different image sets without the need to change the system parameters. We applied five sets of 4D DIC microscope images to our system, and the system detected the nuclei in these sets with similar sensitivity and specificity when we used the same parameter values (Table 1). This reduced sensitivity to differences in image quality makes our system applicable to research. We can apply this system to sets of 4D DIC microscope images of mutant C. elegans embryos (see Additional file 2) and embryos in which specific genes are silenced by RNA interference (see Additional file 3).
Another major advantage of our system is the use of object-tracking algorithms to examine all regions with the features of the image texture of the nucleus (i.e., low local image entropy) in a set of 4D DIC microscope images and to select regions that can actually detect nuclei. A DIC image of a C. elegans embryo contains many regions that have similar (smooth) image textures to that of the nucleus but that do not actually correspond to the nucleus, such as the boundaries between cells and the spaces between the embryo and the eggshell (Figure 3). Thus, in addition to image texture, other features of the nucleus are needed to completely distinguish the nucleus. Our previous system used the (round) shape of the nucleus that was not in the process of cell division in addition to the feature of image texture, as quantified by edge detection operators . This previous system could not detect nuclei in the process of cell division. The object-tracking algorithm in our new system uses spatial and temporal information on the nucleus, and this information is independent of the process of cell division. Thus, our system detects all nuclei – whether or not the cell is dividing – at every time point from one- to 24-cell stages. This continuous detection of nuclei is a great help in following the cell division pattern of the embryo.
Our system effectively detected nuclei over a markedly longer developmental period than did the previous system, i.e., from the one- to 24-cell (Tables 2, 3) stages compared with only the two- to eight-cell stages . This extension of the period of effective nuclear detection primarily results from the very high sensitivity of nuclear detection by low-entropy regions before forward and backward trackings (Table 1). The sensitivity and specificity of nuclear detection by these "original" low-entropy regions depend on the parameters used to produce the regions (i.e., window size and entropy threshold): the higher the sensitivity, the lower the specificity. Our system uses a set of values for these parameters that provides very high sensitivity and very low specificity of nuclear detection by the original low-entropy regions (Table 1), because subsequent forward and backward trackings effectively distinguish those regions that actually detect nuclei from those that do not.
The previous system used a two-step strategy similar to ours: i.e., regions that had the image texture of the nucleus were produced using edge detection operators, and from these "likely nuclear" regions, those that actually detected nuclei were selected using the shape of the nucleus. The sensitivity and specificity of nuclear detection by these "original" likely nuclear regions depended on the parameters used to produce the regions. However, the shape-dependent selection of likely nuclear regions was far less effective than the selection of low-entropy regions by forward and backward trackings. Thus, the previous system used a set of parameter values that provided markedly lower sensitivity and markedly higher specificity of nuclear detection by the original likely nuclear regions than by the original low-entropy regions. In the current study, we found very high sensitivity of nuclear detection by the original low-entropy regions up to the 44-cell stage (data not shown). Thus, improvement in the selection of low-entropy regions will further extend the period of effective nuclear detection. We are developing an improved system that uses both a tracking algorithm and the known shape and size of nuclei in non-dividing cells to select low-entropy regions.
Fluorescent labeling of nuclei is a method that has recently been developed for identifying the positions of the nuclei in living C. elegans embryos [24, 25]. With this method, the genetic information of an embryo is artificially modified so that the embryo expresses nuclear protein fused with fluorescent protein, such as histone H2B fused with green fluorescent protein (GFP) ; the embryo is illuminated by excitatory light (e.g., blue or UV light for GFP), and the expressed fusion protein produces light of a specific color (e.g., green for GFP). Because the nuclei are labeled with a specific color, detection of the nuclei is much easier than that using the DIC microscope. However, the development of the embryo expressing the fusion protein may differ from that of the intact embryo because of the presence of GFP or the modification of genetic information [26–28]. Fluorescent labeling can be used to visualize nuclei for a markedly shorter period than with the DIC microscope because of photobleaching: i.e., the intensity of fluorescence of the fusion protein decreases because of exposure of the protein to the excitatory light , although the amount of photobleaching can be reduced by the use of multiphoton fluorescence imaging . In contrast, the DIC microscope can be used to visualize the nuclei of an intact embryo throughout the development of C. elegans. Therefore, to describe the precise position of nuclei in living C. elegans embryos, identification of the position of the nucleus using the DIC microscope seems more suitable than that using fluorescent labeling of nuclei.
A major drawback of our system is the need for manual selection of low-entropy regions at the first and last time points. These manual operations may reduce the objectivity and productivity of our system, because selection is determined by the operator. However, slight differences in manual selection at the first and the last time points does not influence the automated selection of low-entropy regions in between these points, because the automated selections select all regions that overlap with other selected regions in the adjacent focal plane at the same time point or in the same focal plane at the adjacent time point. Thus, usually our system objectively detects nuclei in between the first and last time points. Manual selection of low-entropy regions at the first and the last time points could still reduce the productivity of our system, because these manual selections usually take about 10 min. However, our system still markedly increases the productivity of identification of the positions of the nuclei in C. elegans embryos, because manual selection of low-entropy regions for all time points from the one- to 24-cell stages (56 focal planes × ~120 time points = ~6720 images) takes more than 50 h. Our system needs about 135 min for computation (120 min for the production of low-entropy regions and 15 min for the forward and backward trackings) and 10 min for manual operations to detect all the nuclei in a set of 4D DIC microscope images of a C. elegans embryo recorded from the one- to 24-cell stages. These times for computation and manual operations are acceptable in research. The selection of low-entropy regions at the first and last time points will be automated, most likely by using known properties of nuclei, such as the known shapes and sizes of nuclei in non-dividing cells.
The low-entropy regions before selection by the forward and backward trackings failed to detect nuclei at around the 44-cell stage or later. Because the window size (10 × 10 pixels) and the threshold value (175) used in this experiment appear likely to be optimal for our system, the result indicates that the limit of the nuclear detection system presented here is around the 44-cell stage. We believe that this limit comes from the reduction in size of the cells during embryogenesis. As the size of the cells decreases during embryogenesis, the distance between the nucleus and cell membrane decreases. Usually at around the 44-cell stage, some nuclei are positioned so close to the cortex of the embryo that a 10 × 10 pixel window cannot produce a high-entropy (> 175) boundary between the nucleus and the image background; the texture of the image background is smooth (Figure 2), and thus the local image entropy in the image background is as low as that in the nucleus. In this situation, the low-entropy regions corresponding to the cortically positioned nucleus merge with the low-entropy regions corresponding to the image background. Because our nuclear detection system removes the low-entropy regions corresponding to the image background, the low-entropy regions produced by our system fail to detect the cortically positioned nucleus. To overcome this limitation, modulation of the window size and/or the threshold value depending on the embryonic stage and/or position of the nucleus within the embryo (central or cortical) might be effective. We observed that low-entropy regions produced using a smaller (< 10 × 10 pixel) window size and/or smaller (< 175) threshold value successfully discriminated between such cortically positioned nuclei and the image background in the later stages of embryogenesis.
Our system is applicable to research programs that require high objectivity and/or productivity of identification of the positions of the nuclei in C. elegans embryos. Because the sensitivity and specificity of nuclear detection by our system depend on the thresholds for minimum overlap ratios (T
), the values of these thresholds should be specified when the system is applied to a specific study. We often use T
= 70% and T
= 4%, because sensitivity is often more important than specificity in our research. We applied this system to our automated cell division pattern measurement system for C. elegans embryos; the measurement system was used in our large-scale cell division pattern analysis of gene-knockout C. elegans embryos . The cell division pattern analysis will provide new opportunities for bioinformatics in studies of the development of multicellular organisms . In addition, this system has been used to measure the positions of the male pronucleus (the sperm-derived nucleus) in a very early C. elegans embryo; the measurements were compared with computer simulations to determine the mechanism that specifies the positions of the male pronucleus during the very early period of C. elegans development . To calculate the precise 3D shape and/or position of a nucleus from the low-entropy regions produced by this system, we need to consider the DIC shear angle, because the angle makes a substantial artifact in DIC images . Because of its high objectivity and productivity of measurement, our system will contribute greatly to studies of the development of multicellular organisms.