Automated Annotating Label in Nanotomy
Using ImageJ to Quantify Nanoparticles in Large EM Datasets
- Fig. 1: Golddigger in action. In pancreas quantum dots (QDs; blue) are targeted to a primary antibody detecting insulin. In addition, immunogold (green) detects G quadruplexes in DNA. This is a reconstruction of 22 Golddigger annotated tiles of 16x16k pixels each. Inset - part of the original image with gold in the nucleus (top) and QDs in insulin-containing vesicles (bottom). Note that in high zoom insulin label is readily distinguishable by eye. At low zoom the vesicles stand out in blue, but also other cells are seen, especially at the top. Bar: 10 μm.
- Fig. 2: Golddigger steps step by step explained. Selected areas of Golddigger used on a single 16k x 16k tile are shown. (a) Original (b) Golddigger main interface (c) result of particle detection as seen in ImageJ (d) larger overview of a part of the dataset, showing labels and in yellow user selected regions of interest and the amount of gold particles (red) in these regions. (e) same view as a and c, now as it is shown in PhotoShop. Note that the amount of label is practically impossible to count manual. In total 25.845 spots were assigned as QDs (blue), 4.192 as gold (red), leading to a total of over 30.037 annotations.
In large-scale electron microscopy (or nanotomy) typically areas of up to 1 mm 2 are recorded with nm resolution. Users thus have a complete overview of a sample at high resolution and cellular context (nanotomy.org). However, at low zoom levels nanoparticles used for label, like gold and quantum dots, are invisible. Golddigger, a Fiji application, detects and marks individual nanoparticles, allowing quantification and visualization of label at all zoom-levels in nanotomy datasets.
Large-scale transmission- (TEM), scanning transmission- (STEM) and scanning (SEM) electron microscopy techniques recently have been developed to make high resolution images of large areas, typically up to ~1 mm2 [1-5] and references therein. The acquired data reveals both context as well as high resolving power, which we therefore name “nanotomy” (for nano-anatomy; reviewed in de Boer et al. ; www.nanotomy.org). To detect specific targets, nanotomy is also being applied on immuno-labelled samples [4,6,7]. Unfortunately electron dense gold or quantum dot (QD) labels are only visible when zoomed in on a small area, whereas examining the tissue context and overall distribution needs analysis of a large field of view. Within a single large-scale acquisition, millions of gold particles may label structures, precluding an analysis of label throughout the dataset. On the other hand, when only a handful of label is present in the dataset, automation will help to recognize these sparse immunolabels. Specific approaches have been developed to recognize gold in some datasets, see for example [4,8-14] and references therein, but these did not fulfill our criteria. To detect different sizes, shapes and opacities of labels in large EM datasets (fig. 1) we use a FIJI  macro that we termed “Golddigger”.
To mark and quantify gold-grains and/or quantum dots in EM images (fig.1 insert; fig. 2A) we considered and/or used algorithms explored by others [4,8-14]. We set out to use Difference of Gaussian (DOG)-, Hessian- and Laplace filtering, amongst others.
A Laplace-like filtering meets our requirements at best. When optimized for microscope settings and properties of the nanoparticles used for labelling, Golddigger can be directly run to mark and quantify label. We note that the size at full resolution (>3 Gb in fig. 1) of large-scale images precludes processing in ImageJ. Therefore, Golddigger calculations are performed on individual tiles prior to stitching. The workflow of label detection is summarized as follows: A copy of the original image is made, and added in a second layer. If necessary a gamma correction is applied to allow better detection of the particles and the image is converted to 32-bit floating followed by a Gaussian Blur with a sigma of 1. The outcome of this image is convolved with a 3x3 (0,1,0 / 1,-4,1 /0,1,0) kernel to highlight small clusters of bright pixels. An automatic threshold is applied (fig. 2B) to the resulting image using the Renyi Entropy dark function of ImageJ. If separation of particles needs to be improved, a Watershed algorithm can be applied, but this slows down the process significantly. Next, ImageJ’s Analyze Particles function detects all dots of a certain size and circularity. The positions of these are stored in the ROI manager and transferred to the original image. On this image the positions are one- by one measured and checked if they meet the eight user set conditions to define the particles. These conditions are based on intensity (1: mean; 2: maximum; 3: minimum; 4: standard deviation high; 5: standard deviation low), size and shape (6: area; 7:roundness; 8: skew). If not all user criteria of a particle are met, these should be dismissed and the position is stored to layer 1 in the ROI manager. All other particles do fulfil the criteria and will be stored to the second layer. Both the negative and positive labels as well as the measurements results remain available during the whole procedure. The user evaluates the outcome and corrects the eight criteria above to include more true labels or exclude false-positives if needed. This evaluation is based on the values stored in the table and therefore does not require additional image-analysis and therefore is fast and can be repeated iterative. If satisfactory, Golddigger will redo the detection using the latest parameters which are automatically stored in a text file and as metadata in the image file, and negative hits are deleted. The final export is a tiff image, which includes all positively detected grains that full-fill the criteria (i) encircled in an ImageJ overlay (fig. 2C). These marking circles are scaled automatically and thus visible at any zoom level; or (ii) coded with a dot instead of a circle in an ImageJ overlay to be used in the incorporated grain counting function (fig. 2 D,E); or (iii) as circles only at grain positions useable as an overlay in Photoshop (not shown). Settings for one tile are then applied to the whole series. This allows quantitation of label in each tile (fig. 2), or an overview of all label in the sample when stitched (fig. 1). The user manual available as supplementary data, step-by-step describes Golddigger in detail.
Concluding Remarks & Future Considerations
Golddigger combines characteristics of label to help to identify particles used in immuno-EM detection of targets. In EM images the black particles are identified against the ultrastructural background, note that contrast differences between label and ultrastructure carefully chosen during acquisition will help to identify the particles. Inherent to immunolabeling is that, at the level of epitope/ antibody interaction, false-positives and false negatives always will be present . In immuno-EM our estimate is that labeling of only a few percent of target molecules can be typically achieved . The odds ratio of the technique is likely much higher than the mistakes regarding false-positives and false-negatives made by the current version of Golddigger. The use of iterative cycles specific for the dataset to correct for inclusion of false-positives and exclusion of false negatives may be further improved. Here we show that particles of different density and shapes, with similar sizes, can be discriminated which allows analysis of multi-labeling in EM. Given the amount of EM-data recorded with automated acquisition and the need for high zoom to see that nanoparticles, detecting these labels in large datasets is manually, i.e. screening with the human eye, an exercise that practically would take too long. The combination of nanotomy with Golddigger will help detection and annotation of label, including colocalization, thus aid in multi-scale interpretation and quantitative analysis in large-scale electron microscopy.
Rat pancreas was immunolabeled and recorded with STEM as described [1, 7].
We thank J. Kuipers and P. de Boer for data and discussions, and acknowledge financial support from the Netherlands Organization for Scientific Research.
A manual (GolddiggerManual) and the ImageJ files Golddigger.ijm and Golddigger Tools.ijm are available via nanotomy. org.
 Hoffmann, R.F., Moshkin, Y.M., Mouton, S., Grzeschik, N.A., Kalicharan, R.D., Kuipers, J., Wolters, A.H., Nishida, K., Romashchenko, A.V., Postberg, J. et al. (2016). Guanine quadruplex structures localize to heterochromatin. Nucleic Acids Res. 44:152-163. DOI: 10.1093/nar/gkv900
 Sokol, E., Kramer, D., Diercks, G.F., Kuipers, J., Jonkman, M.F., Pas, H.H., and Giepmans, B.N. (2015). Large-Scale Electron Microscopy Maps of Patient Skin and Mucosa Provide Insight into Pathogenesis of Blistering Diseases. J. Invest. Dermatol. 7, 1763-1770. DOI: 10.1038/jid.2015.109
 Kuwajima, M., Mendenhall, J.M., and Harris, K.M. (2013). Large-volume reconstruction of brain tissue from high-resolution serial section images acquired by SEM-based scanning transmission electron microscopy. Methods Mol. Biol. 253-273. DOI: 10.1007/978-1-62703-137-0_15
 Faas, F.G., Avramut, M.C., van den Berg, B.M., Mommaas, A.M., Koster, A.J., and Ravelli, R.B. (2012). Virtual nanoscopy: generation of ultra-large high resolution electron microscopy maps. J. Cell Biol. 3, 457-469. DOI: 10.1083/jcb.201201140
 Ravelli, R.B., Kalicharan, R.D., Avramut, M.C., Sjollema, K.A., Pronk, J.W., Dijk, F., Koster, A.J., Visser, J.T., Faas, F.G., and Giepmans, B.N. (2013). Destruction of tissue, cells and organelles in type 1 diabetic rats presented at macromolecular resolution. Scientific reports 3:1804. DOI: 10.1038/srep01804.
 de Boer, P., Hoogenboom, J.P., and Giepmans, B.N. (2015). Correlated light and electron microscopy: ultrastructure lights up! Nat. Methods 6, 503-513. DOI: 10.1038/nmeth.3400.
 Kuipers, J., de Boer, P., and Giepmans, B.N. (2015). Scanning EM of non-heavy metal stained biosamples: Large-field of view, high contrast and highly efficient immunolabeling. Exp. Cell Res. 2, 202-207. DOI: 10.1016/j.yexcr.2015.07.012.
. Płaczek, B., Bułdak, R.J., Polaniak, R., Matysiak, N., Mielanczyk, Ł, and Wojnicz, R. (2014). Detection of immunogold markers in images obtained from transmission electron microscopy. Journal of Medical Informatics & Technologies 111--117.
 Płaczek, B., Bułdak, R.J., Brenk, A., and Polaniak, R. (2015). Automatic Evaluation of Area-Related Immunogold Particles Density in Transmission Electron Micrographs. In Intelligent Information and Database Systems, Springer) pp. 53-61.
 Mayhew, T.M. (2011). Quantifying immunogold localization on electron microscopic thin sections: a compendium of new approaches for plant cell biologists. J. Exp. Bot. 12, 4101-4113. DOI: 10.1093/jxb/err176.
 Wang, R., Pokhariya, H., McKenna, S.J., and Lucocq, J. (2011). Recognition of immunogold markers in electron micrographs. J. Struct. Biol. 2, 151-158. DOI: 10.1016/j.jsb.2011.07.005
 Monteiro-Leal, L.H., Troster, H., Campanati, L., Spring, H., and Trendelenburg, M.F. (2003). Gold finder: a computer method for fast automatic double gold labeling detection, counting, and color overlay in electron microscopic images. J. Struct. Biol. 3, 228-239.
 Rolland, J.P., Bon, P., and Thomas, D. (1997). MACS: automatic counting of objects based on shape recognition. Comput. Appl. Biosci. 5, 563-564. DOI: 10.1093/bioinformatics/13.5.563
 Mayhew, T.M., Muhlfeld, C., Vanhecke, D., and Ochs, M. (2009). A review of recent methods for efficiently quantifying immunogold and other nanoparticles using TEM sections through cells, tissues and organs. Ann. Anat. 2, 153-170. DOI: 10.1016/j.aanat.2008.11.001.
 Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T., Preibisch, S., Rueden, C., Saalfeld, S., Schmid, B. et al. (2012). Fiji: an open-source platform for biological-image analysis. Nat. Methods 7, 676-682. DOI: 10.1038/nmeth.2019.
 Schnell, U., Dijk, F., Sjollema, K.A., and Giepmans, B.N. (2012). Immunolabeling artifacts and the need for live-cell imaging. Nature methods 2, 152-158. DOI: 10.1038/nmeth.1855.