CAS: Cell Annotation Software – Research on Neuronal Tissue Has Never Been so Transparent
CAS (Cell Annotation Software) is a novel tool for analysis of microscopic images and selection of the cell soma or nucleus, depending on the research objectives in medicine, biology, bioinformatics, etc. It replaces time-consuming and tiresome manual analysis of single images not only with automatic methods for object segmentation based on the Statistical Dominance Algorithm, but also semi-automatic tools for object selection within a marked region of interest. For each image, a broad set of object parameters is computed, including shape features and optical and topographic characteristics, thus giving additional insight into data. Our solution for cell detection and analysis has been verified by microscopic data and its application in the annotation of the lateral geniculate nucleus has been examined in a case study.
KeywordsNeuroscientific data Image processing Image understanding Neuron segmentation Nuclei segmentation Cell segmentation Microscopic images Statistical Dominance Algorithm Histological images Immunohistological images Fluorescent microscopy 2D microscopy Shape features
Studying the principles of the structure of neuronal nets can bring us closer to unraveling the functioning of heterogenic populations of neurons within the central nervous system. The investigation of these nets involves consideration of the characteristics of neurons and their distribution within a particular nervous structure. In some cases, the distribution of neurons is enough to predict their characteristics. Often neurons form repetitive, regularly organized groups within which they share common physiological and neurochemical characteristics. For example, this phenomenon is pronounced within the visual cortex (Dyck & Cynader 1993; Kaas 2012; Ohki et al. 2005) and the spinal cord (Merkulyeva et al. 2016). However, in many other cases the structure of the object of interest is not so highly ordered; however, this does not mean that neurons are randomly distributed. A good example is a laminated visual thalamic structure, the lateral geniculate nucleus, LGN, which contains neurons differentiated by multiple features, such as input from ipsilateral or contralateral eye, the visuotopic position of retinal projections, or the type of projecting ganglion cell (Sanderson 1971; Payne & Peters 2002; Schiller 2010). In such cases, we expect these neuronal nets to have a complex structure; a full description of them can be obtained only by means of thorough analysis of cell parameters measured within the entire structure. Obviously, this leads to a significant increase in the effort needed to detect and analyze neuronal data, but this can be significantly facilitated with the use of specialized software.
The software applications designed for this purpose offer a wealth of opportunities for manual cell detection; however, in many cases only restricted opportunities for automatic cell detection are provided, usually based on brightness and color filters. Sometimes these techniques are suitable, but there are many cases in which biological objects have staining that is too difficult to be entrusted to automatic detection modes. Examples include a high level of noise, such as stained fibers and dendritic arbors, or high variability of the color intensity of objects of interest. Moreover, subsequent analysis is performed in statistical data processing software such as Statistica (StatSoft) and SPSS Statistics (IBM). These applications have a specific way of organizing data, but image processing software cannot export data to these programs; therefore, a lot of additional work is necessary for data preparation.
Cell Annotation Software, CAS1, was created in order to meet the needs arising from investigations of objects in histological specimens. ’Object’ is a general term we use throughout whole paper, but it should be understood as cell soma or cell nucleus, since either can be selected with CAS depending on the chosen parameters. The selection procedure ensures that any artifacts that are a result of tissue processing are considered. Comparative analysis of neuronal populations is also possible. Moreover, easy navigation between various levels of analysis and final data storage in files whose format is easily accessible to other software makes CAS a unique and fully functional program.
The program is designed to work with histologically and immunohistologically prepared specimens of neuronal tissue for which topological information about neural object placement is crucial. It allows different parts of stratified structures (e.g. laminae of neocortex, thalamic nuclei, spinal cord) to be compared and their intrinsic inhomogeneity to be examined. Definition of a region of interest of any shape is very practical and easy to apply. The semi-automatic detection of objects depends on only a few parameters; the rules of how to set these parameters are given in this work. Data unification in the X-axis allows for better comparison between several specimens; this is important in ontogenetic research, in which the absolute size of brain structures changes progressively.
Brightness normalization is important for quantitative comparison of the degree of histochemical or immunohistochemical staining of tissues. This makes it possible to analyze the amount of stained substance in objects indirectly, which in turn allows examination of the accumulation (or degradation) of a substance during maturation of the nervous system or under experimental conditions
Combining the aforementioned functionality of this tool makes CAS a powerful solution which facilitates histological structure analysis and removes the manual work which is normally necessary with other open-source software.
The basic discussion of selecting the most accurate image processing algorithms for object detection is presented in “Object Segmentation”. This draws attention to many problems which must be considered in order to choose a suitable method for data segmentation. Detecting objects requires automatic description of shape, size, and distribution. The implemented measures and their interpretations are given in “Shape Parameters”. Finally, Section “Case Study” presents an example of the application of CAS in research on the LGN. The “Conclusions” summarize the article.
Although it has already been addressed, the problem of correct segmentation between cell soma/nucleus and backgrounds in extensive groups of microscopic images does not yet seem to have been universally solved. This is interesting because the problem of object detection in biplanar images (in which there are only objects and a well-defined background) is widely discussed in the literature and there are many methods with various degrees of accuracy for solving it (Irshad et al. 2014).
The standard approach is based on a family of binarization methods and other more advanced algorithms dedicated to this problem. In most cases, researchers conduct object segmentation manually, exploiting the tool-set in ImageJ/Fiji (Papadopulos et al. 2007; Collins & et al. 2007; Schneider et al. 2012; Schindelin et al. 2012; Hartig 2001). Researchers, who specialize in image processing, have also created solutions dedicated to specific applications. Examples include the ImageJ plug-in (Forero et al. 2010; Pool et al. 2008), an ImageJ framework expansion (Gulyás et al. 2016), software for the Matlab environment (La Torre et al. 2013), and stand-alone applications such as CellProfiler (Kamentsky et al. 2011; Carpenter et al. 2006). Commercial solutions also exist, but their license usually precludes them from comparison. In many of the aforementioned solutions, extended H-minima is exploited for segmentation, while in our work the Statistical Dominance Algorithm SDA (Piorkowski 2016) is applied; this is novel in this field.
This section presents the common problems which should be considered when object segmentation is performed on microscopic images. The discussion starts with the choice of color space and its influence on the accuracy of the location of the border of the final object. The difficulties which must be overcome by binarization procedures are then described. Finally, a comparison of the standard procedures exploiting the extended H-minima method and the novel approach based on the SDA algorithm applied in the presented software is given.
Adjusting Color space
Data gathered by microscopic examination is stored as color images. Generally, the correct approach to further analysis of color data assumes the necessity of color-space normalization (Ing et al. 2016). However, in the case of microscopic nerve tissue data, which is the main objective of CAS development, normalization is not necessary due to the monochromatic nature of the input data, which holds most information in one or two channels of the three used for acquisition (usually in RGB format). Moreover, the differences between channels are not substantial. Previous attempts at selecting the most representative color channel of various tissues have shown that the best solution is to choose the channel with the highest variance (in RGB) for processing (Kolodziejczyk et al. 2014).
Standard approach to image binarization
The standard approach to object segmentation has already been established and is based on microscopic preparations, a commonly used solution that was introduced in neuro-science (Selinummi et al. 2005; Baecker 2010; Hartig 2001). This technique consists of two stages of image processing: background correction and application of one of a range of binarization methods.
Background correction aims to remove the very common phenomenon of uneven illumination. This can be achieved in various ways, for instance by high-pass filtering (which removes low frequencies). However, the rolling ball algorithm, which is the most widely applied, is based on background subtraction, as suggested by Sternberg (1983). This procedure subtracts average illuminance in a usually disc-shaped given neighborhood and prompts the operator to specify its dimensions. It is assumed that this size (the radius) should be a few times larger than the object’s area in order to avoid distorting objects while maintaining the effective removal of uneven illumination. A discussion on the influence of changing radius on the performance of the rolling ball algorithm was presented by Ekstrom et al. (2016).
Binarization is the second stage; it assigns 1s for pixels belonging to the object and 0s for the background. The easiest approach uses global thresholding, which defines one threshold value for the whole image. This is an efficient solution for a group of selected images; however, when larger variations in image content are considered or noise and distortions are present, more sophisticated approaches are necessary. In such cases, adaptive methods are useful, such as the essential method introduced by Otsu (1975). ImageJ/Fiji software (Hartig 2001) also implements other methods which can be exploited (e.g. Huang, Li, Max Entropy, Renyi Entropy, Shanbhag, Yen, etc.). The common advantage of all adaptive methods is the lack of necessity for the operator to set parameters, hence the human influence on data processing is removed. Still, these techniques may misinterpret areas of images that have different object density and background type than the rest of the image, or where images containing tissues have varying structures.
Extended H-minima Technique
Statistical Dominance Algorithm
We suggest a novel approach for object segmentation based on the recently introduced Statistical Dominance Algorithm (SDA) Piorkowski (2016), which exploits statistics to specify image content. For each pixel in a monochromatic image, a number of neighboring pixels are defined such that the currently analyzed pixel is dominant over them (or inversely, neighboring pixels are dominant over the current pixel). As with H-minima, it is possible to define an additional threshold that must be reached to assume dominance. Thus, the output image contains statistics of point distribution in the input data; however, it is interesting that this method makes it possible to keep the original shapes. Because relations between points instead of illumination differences are computed, sensitivity to uneven illumination and structure variation diminishes. Moreover, this technique can derive indirect information about an object’s area.
imgin – input image,
imgout – output image,
SX, SY – width and height of input/output image,
R – radius of neighborhood,
N – size of neighborhood mask (also size of mirror margin used here), N = ⌈R⌉
T - the threshold to be checked (especially for noisy images); its value usually is positive.
in the upper left: fragment of larger, uniform tissue (or aggregation of objects);
in the lower left corner: regular cell, with lighter luminance;
in the upper right corner: cell with dendrites, with a darker luminance;
in the bottom right corner: background luminance unevenness (levels 60–70–80).
objects of a small, round shape, regardless of real brightness, are clearly promoted if only the minimum threshold condition is met;
both cells have comparable statistical values despite the difference in brightness in the input image; the normalization of these values is shown in the real microscopic image (Fig. 3b);
dendrites have the highest statistical values;
The above considerations show that the characteristics of the SDA algorithm justify its use in the analysis of nerve cells. Preliminary studies have shown the utility of segmentation algorithms in other tissues and histopathological images for which classical methods such as Otsu or adaptive thresholding do not provide accurate solutions (Kowal et al. 2013).
As described, it is necessary to set two parameters: the size of the neighborhood, specified by radius R (which defines a circular neighborhood), and minimal difference threshold T, which is analogous to the binarization threshold H exploited in H-minima. In the case of the neighborhood radius, the procedure is similar to the one used in background subtraction: usually a radius two or three times longer than the object is sufficient and testing for larger values does not change the outcome (this is presented more precisely in “Comparison of Object Detection Methods”). The selection of thresholding value, as in H-minima, also demands operator interaction. However, some datasets containing large groups of images give similar, stable results for the same parameter settings (a group of images accessible at BrainMaps (2005) was used for verification). The novel approach implemented in CAS assumes exploitation of the SDA algorithm in the background subtraction stage, followed by manual or automatic application of global or adaptive binarization. The image obtained from SDA should also be binarized, which can be done automatically (in the case of CAS, the Otsu method is suggested as it can remove some types of disturbance; refer to Fig. 4), or the global threshold can be set manually. The presented research focuses on object detection with the application of SDA, whose compatibility with other techniques has been proved.
This image is too difficult for the standard object segmentation method, but was successfully segmented by applying SDA. The complex input data (see Fig. 5a) shows sagittal sections of cerebellum cut on a Leica microtome and stained with cresyl violet. The greatest challenge here is the presence of different tissues in the image. The standard approach based on binarization does not correctly segment the cells from the (darker) tissue – the granular layer – regardless of whether the manual or automatic global thresholding method implemented in ImageJ/Fiji is used. An example of the outcome is depicted in Fig. 5b. Since the SDA’s output gives greyscale values (see light-grey and dark-grey objects in Fig. 5c), it was possible to continue with further segmentation and select an appropriate threshold in order to obtain accurate cell detection, as annotated on the original image in Fig. 5c. This feature proves the attractiveness of the SDA method for semi-automatic segmentation of cells in nerve tissues (Kopacz & Piorkowski 2017). The presented solution does not perform automatic segmentation of clustered nuclei; although this problem is still unresolved, some approaches are addressed by (Koyuncu et al. 2016; Jung & Kim 2010; Cheng et al. 2009; Gertych et al. 2012; Skobel et al. 2017).
Comparison of Object Detection Methods
In order to choose a method for further processing, the agreement in object detection in biplanar microscopic images was evaluated. Good quality resolution of objects and backgrounds was assured in all examples. For the standard approach, the global threshold was selected manually to assure correct segmentation. The radius applied in background subtraction methods and SDA was set to 50 pixels.
We then concentrated on the relation between SDA parameters applied for images of objects of varying sizes. The influence of varying radius on detected object area is presented in Fig. 6d for the image presented in Fig. 6c, which was taken from a public dataset prepared for the Nucleus Counting Challenge competition that took place at the 2015 BioImage Informatics Conference. From this plot, one can see that the outcome stabilizes for radii above 20 pixels; the bigger the detected threshold value, the smaller the influence of data size. Of course, when the image is resized, the radius should be scaled as well. Moreover, the radius is always related to the object size, but the relation to the threshold value is nevertheless maintained.
The presented experiments and considerations address all the problems encountered in achieving accurate object segmentation in microscopic images of various tissues. From the elaborated results, it was decided that the CAS tool should interpret input images using only one data channel. The SDA was suggested as a segmentation method; however, to maintain compatibility with other research tools, the extended H-minima is also accessible. Finally, setting method parameters demands knowledge about object size, which for method radius is straightforward, while the SDA threshold should be adjusted and the final binarisation threshold can be chosen automatically using the Otsu approach.
It is not sufficient to develop techniques for automatic detection of objects in images: in most tasks, especially when general insight into object population is sought, this is the starting point for further analysis which demands thorough inspection of object features. Depending on the requirements, the desired characteristics vary, yet in general they concentrate on basic shape features which allow analysis of not only circularity or elongation, but also spatial and size distribution within a sample. This area of research is very tiresome, therefore automation is of great importance. Below, several shape parameters implemented in CAS are detailed and an interpretation is given with some examples.
General features computed for exemplary neural cell presented in Fig. 8
Topological features computed for different cells
A [ μm2]
P [ μm]
It is difficult to find a pattern in which C4, C5, and Roughness are considered. The results gathered for C4 seem completely unrelated to the shape, as one can see that using only the circumference for description is insufficient. A similar observation is also seen for the Roughness parameter, for which only area was exploited in the calculation. However, for C5, the feature value is still strongly related to the object size, yet smaller values represent more circular shapes when objects with comparable area are considered (for instance, check the outcomes in rows 4–6, 7 with 8, etc.).
Next, when the Malinowska parameter is taken into account, values close to 0 describe circular objects, while higher values are related to longer shapes; this is also true of the Shapeless parameter. Moreover, we can additionally state that this works inversely for C1 when the formulas are compared. Finally, Roundness is equal to 1 for circular objects and gets smaller for longer shapes.
The last group of parameters that describe objects should help to organize this data and obtain more spatial information about it. First of all, it is possible to mark a region (bigger patch) on the image and work only with the objects belonging to it. Such a region can be automatically divided into three perpendicular horizontal layers. Next, for each object the major and minor axes are found. Here, two algorithms are exploited:
- Ellipse Fit
– chooses the best fitted ellipse that can be inscribed in the object. The major and minor axes correspond to those of the ellipse. Moreover, the centroid can be recalculated to overlap the ellipse center point.
– a shape feature parameter that computes the relation between the two longest and perpendicular dimensions of an object. Principal component analysis (PCA) is exploited to gather the results. The major axis corresponds to the direction of the first eigenvector; the minor axis is related to the second one.
In both cases, the reference angle is computed as the angle between the object’s major axis and the image’s x-axis. This is used to position objects better in the input data.
Data Processing Example
Histology and Image Acquisition
All experimental procedures were approved by the Ethics Commission of the Pavlov Institute of Physiology. Experiments were performed in strong accordance with the requirements of Council Directive (2010/63/EU) of the European Parliament on the protection of animals used for experimental and other scientific purposes.
monoclonal mouse primary antibodies to NeuN (Millipore, MAB377, in 1:5000 dilution);
monoclonal mouse primary antibodies to non-phosphorylated Neurofilament H (NF-H) (Covance, SMI-32, in 1:3000 dilution);
polyclonal rabbit primary antibodies to the parvalbumin (Abcam, ab11427, in 1:10000 dilution).
Biotinylated secondary antibodies (goat anti-rabbit IgG, BA-1000, or horse anti-mouse IgG, BA-9200; Vector Laboratories, Peterborough, United Kingdom) for further DAB/Ni staining protocol for light microscopy.
Fluorochrome-conjugated secondary antibodies (Alexa Fluor488 goat anti-mouse IgG or Alexa Fluor568 goat anti-rabbit IgG, Thermo Fisher Scientific, Waltham, MA, USA) for cell detection with a fluorescence microscope.
DAB-reacted slices were analyzed with an Olympus microscope (Olympus Corporation, Tokyo, Japan, a 10 × magnification lens) using a Nikon photo camera (Nikon Corporation, Tokyo, Japan). Fluorescent slices were analyzed with a Leica DMI6000 inverted fluorescence microscope (at the Center for Molecular and Cell Technologies, Research Park, St. Petersburg State University).
Mark the layer borders.
Mark the sub-layer parts of layers.
Outline all labelled cells with appropriate parameters one-by-one.
Measure data and copy the results to an external file.
Complete the measurement file with additional information about the layer, sublayer, cell location, etc.
Perform cell distribution normalization on the X axis.
Perform cell brightness normalization using a reference region.
Description how data preparation staps are performed with and without CAS
Cell annotation software
Semi-manual (needs parameter setting)
Automatic (allows parameter selection)
Automatic (uses data generated by user)
Automatic (user sets the normalization coefficients)
Automatic (user sets the reference region)
Comparison of performance when manual outlining and semi-automated solution accessible in CAS were exploited
Cell segmentation speed-up
Cell segmentation with corrections speed-up
In addition, the software has convenient instruments for data presentation and re-examination. It is also worth pointing out that if changes are made (e.g. removing or adding cells, etc.), the results are recomputed automatically without operator intervention, which makes corrections and changes easy to apply. As a result, we have a convenient tool to detect specific changes of the LGN under experimental conditions or as a consequence of age-related development.
This work addresses the problem of automatic cell soma or nucleus segmentation and annotation in microscopic images. A detailed comparison of the standard image processing method against a novel approach exploiting SDA was presented. The research shows that the method used in CAS enables object border detection with comparable quality to standard methods. In difficult cases in which there are more types of object to be recognized, it outperforms the current leading techniques. In addition, shape features were introduced for description of object characteristics and specific analysis and interpretation were performed. Additional features implemented in CAS which enable data scale changes, data division into layers, and other functionality were presented with a case study of semi-automatic annotation of the LGN of cats.
Cell Annotation Software was designed to meet the need for a system which combines the various functionalities necessary for segmentation and analysis of objects in neural specimens. The presented approach proved to fulfill all demands and allows: saving topological information about cell distribution by region definition; comparative analysis of different slices and investigation of changes in the neuronal population due to x-axis normalization; and quantitative comparison of the degree of histological or immunohistological staining of tissues in slices when a change in the production quantity of a particular substance due to the influence of experimental factors is investigated. Moreover, simple recalculation of results and storage in open-format files make this software easy to incorporate in research pipelines.
The practical application of CAS in research conducted at the Pavlov Institute of Physiology proved to be a great culmination of this research, but a new development frontier has also arisen. In future work, we plan to address the problem of multiple labeling, working with tissues with various cell types and 3D reconstruction.
Information Sharing Statement
The presented Cell Annotation Software can be downloaded from http://home.agh.edu.pl/pioro/cas/.
The authors thank Shkorbatova P. for NeuN primary antibody, prof. Musienko P. for microscopic equipment, and the Center for Molecular and Cell Technologies, Research Park, St. Petersburg State University.
K. Nurzynska was supported by statutory funds for young researchers (BKM/507/RAU2/2016) from the Institute of Informatics, Silesian University of Technology, Poland.
A. Piorkowski was supported by AGH University of Science and Technology, Faculty of Geology, Geophysics and Environmental Protection as a part of a statutory project.
- Baecker, V. (2010). Workshop: Image processing and analysis with ImageJ and MRI Cell Image Analyzer.Google Scholar
- BrainMaps (2005). BrainMaps: An Interactive Multiresolution Brain Atlas. http://brainmaps.org.
- Carpenter, A. E., Jones, T. R., Lamprecht, M. R., Clarke, C., Kang, I. H., Friman, O., Guertin, D. A., Chang, J. H., Lindquist, R. A., Moffat, J., et al. (2006). Cellprofiler: image analysis software for identifying and quantifying cell phenotypes. Genome Biology, 7(10), R100.Google Scholar
- Dyck, R. H., & Cynader, M. S. (1993). An interdigitated columnar mosaic of cytochrome oxidase, zinc, and neurotransmitter-related molecules in cat and monkey visual cortex. Proceedings of the National Academy of Sciences of the United States of America, 90(19), 9066–9069.CrossRefPubMedPubMedCentralGoogle Scholar
- Ekstrom, A., Suvanto, R. W., Yang, T., Ye, B., & Zhou, J. (2016). Robust neuron counting based on fusion of shape map and multi-cue learning. In International conference on brain and health informatics (pp. 313). Springer.Google Scholar
- Gonzalez, R. C., & Woods, R. E. (2001). Digital image processing, 2nd Edn. MA, USA: Addison-Wesley Longman Publishing Co., Inc.Google Scholar
- Hartig, S. M. (2001). Basic image analysis and manipulation in ImageJ. In Current protocols in molecular biology. Wiley. ISBN 9780471142720. https://doi.org/10.1002/0471142727.mb1415s102.
- Ing, N., Salman, S., Ma, Z., Walts, A., Knudsen, B., & Gertych, A. (2016). Machine learning can reliably distinguish histological patterns of micropapillary and solid lung adenocarcinomas. In Information technologies in medicine, advances in intelligent systems and computing (Vol. 471, pp. 193–206). Springer.Google Scholar
- Kamentsky, L., Jones, T. R., Fraser, A., Bray, M. A., Logan, D. J., Madden, K. L., Ljosa, V., Rueden, C., Eliceiri, K. W., & Carpenter, A. E. (2011). Improved structure, function and compatibility for cellprofiler: Modular high-throughput image analysis software. Bioinformatics, 27(8), 1179–1180.CrossRefPubMedPubMedCentralGoogle Scholar
- Kaplan, E. (2014). The new visual neurosciences. In Werner, J.S., & Chalupa, L.M. (Eds.) The M, P, and K pathways of the primate visual system revisited (pp. 215–227). Cambridge: MIT Press.Google Scholar
- Kolodziejczyk, A., Ladniak, M., & Piorkowski, A. (2014). Constructing software for analysis of neuron, glial and endothelial cell numbers and density in histological nissl-stained rodent brain tissue. Journal of Medical Informatics and Technologies, 23, 77–86.Google Scholar
- Kopacz, B., & Piorkowski, A. (2017). A Review of the Efficient Algorithm Implementation for Image Processing in the ImageJ and Matlab Environments (Przeglad moŻliwości wydajnej implementacji algorytmów przetwarzania obrazów w środowisku Matlab). Studia Informatica, 38(132), 177–190.Google Scholar
- La Torre, A., Alonso Nanclares, L., Muelas Pascual, S., Peña Sanchez, J. M., Oroquieta, F., & et al. (2013). 3d segmentations of neuronal nuclei from confocal microscope image stacks. Frontiers in neuroanatomy, 7, 1–10.Google Scholar
- Maric, D. (2015). 2015 BioImage Informatics Conference: Nucleus Counting Challenge. https://isg.nist.gov/BII_2015/webPages/pages/nucleusCounting/NucleusCounting.html.
- Otsu, N. (1975). A threshold selection method from gray-level histograms. Automatica, 11(285–296), 23–27.Google Scholar
- Payne, B. R., & Peters, A. (2002). The cat primary visual cortex. In Payne, B.R. & Peters, A. (Eds.) The concept of cat primary visual cortex (pp. 1–129). San Diego, London, Boston, New York, Sydney, Tokyo, Toronto: Academic Press.Google Scholar
- Piorkowski, A. (2016). A statistical dominance algorithm for edge detection and segmentation of medical images. In Information technologies in medicine, advances in intelligent systems and computing (Vol. 471, pp. 3–14). Springer.Google Scholar
- Russ, J. (1998). The image processing handbook, 3rd Edn. Springer, and IEEE Press: CRC Press.Google Scholar
- Skobel, M., Kowal, M., & Korbicz, J. (2017). Nuclei recognition using iterated conditional modes approach. In International conference on computer recognition systems (pp. 326–335). Springer.Google Scholar
- Stichel, C. C., Singer, W., & Heizmann, C. W. (1988). Light and electron microscopic immunocytochemical localization of parvalbumin in the dorsal lateral geniculate nucleus of the cat: evidence for coexistence with GABA. The Journal of Comparative Neurology, 268(1), 29–37.CrossRefPubMedGoogle Scholar