you want to count the number of black parts? some black parts may connect to each other and as a result the particle analyzer take them as one particle......... in this case the size of particles should be involved.

Hi Jeannie,

The Analyze Particles... command in ImageJ can sometimes be a bit confusing, and you need to look at the results very carefully to make sure it is doing what you think it is doing.

I suspect that in your case ImageJ is measuring the *white* regions and not the black regions. To be sure, choose 'Image -> Adjust -> Threshold...'. If you have got the option 'Red' below the histogram, then the region that the threshold command turns red is the region that will be analysed. If that's not what you want, click on the 'Dark background' checkbox to change it. I like to always set a threshold like this before running Analyze Particles... since otherwise sometimes ImageJ will measure the black regions and sometimes the white regions*...

Some other Analyze Particles tips:

- Consider using 'Add to Manager' to get a list of individual ROIs (regions of interest) for each particle. Then you can click on each ROI and examine it individually. (However note that these ROIs only give the outer border of every particle... which can include any holes inside).

- If you are using 'Display results', it is usually best to choose 'Clear results'. Otherwise you risk having both old and new measurements in your results table. What makes this particularly scary is that this can happen even if your results table wasn't visible at all... i.e. if you run Analyze Particles... without 'Display results', then it can still be merrily adding measurements to the results table, which will only appear the next time some command shows the results table. So choose 'Clear results' to make sure all the measurements are those made by Analyze Particles and nothing else. If you need to keep an old results table along with a new one, click on it and choose 'File -> Duplicate...' first.

- If you choose 'Show: Count Masks' then this creates a new image in which each counted region has got a different numeric value. This can also help you see what Analyze Particles was measuring.

*- I think it depends on what lookup table you are using, and your settings under Process -> Binary -> Options. But rather than worry about this, setting a threshold and seeing what turns red is a much easier visual check.

I agree with Hossein. Particle analyzer uses thresholding to segment. This leads to many of your particles joining together. Is this image binaries from a threshold, or is this how it looks originally?

One option is to auto threshold (your choice of desired algorithm) then after converting to binary, use the watershed feature to break apart joined particles. This works to some degree but probably not quite as you want. Another option, if you're only concerned with distribution, is to find local minimas and measure the centroid x,y location. You could input these locations into R and use the spatstat package to do a complete spatial analysis on your dataset.

Hi, Kristopher,

Yup, this image is binarized from a threshold.

Comparing to the original image, though not all details can be highlighted (e.g. corners with large tortuosity and very small particles), the particles can generally picked out, whose number suppose to be more than I got from the Particle Analyzer.

For the alternative option you gave, it seems not feasible in my case, cause there are many particles need to be processed.

Hi, Peter,

Yup, the issue about which part is being analyzed is truly confusing in ImageJ. I followed your suggestion, went through the threshold and Analyzer again, from the 'mask count', this time the results seems better, but I still found some problem there.

I attached the 'mask count' picture, in the red circle, the inner edge of the '352' particles was not sketched out, does that mean that the size of the particles was over-estimated?

For the case in which size of particles are small and largely dispersed, the overestimation seem not be a big issue. Then I apply the same procedure to a sample with larger size particles dispersed, secondly attached, then in the mask count counterpart, you will see that in the red circle, considerable mount of edge were not detected. Does that greatly deviate the result?

Post the original image and we might be able to work up an approach for you

Using Fiji distribution of ImageJ I can think of a couple methods using either Laplacian or Canny edges to isolate particle counts and location

sure, here is the original picture.

Ok

Here is a simple method using local minima. It looks to under estimate some of the particle dimensions but you can play with the noise tolerance to get a better fit. The following code can be pasted into the macro editor and run or alternatively I have attached it as a .ijm file.

This is significantly different from your results above. I tried to optimize separation of particles from each other. Only you can tell whether the segmentation is appropriate to your particles or not. If this doesn't work well for you, you could try the first steps through #6 then proceed to general thresholding for Analyze particles.

/* Overview of steps for analysis

* Use steps outlined for manual analysis or run this macro after opening and image

* for an autmated version of the analysis. To customize a macro step enabling macro

* recording (Plugins > Macro > Record...) then perfom the step as desired. Copy the

* macro recorded step and replace the appropriate step in this macro. Use the // to

* disable a specific line.

*

* 1. Open image (File > Open)

* * Beware image titles with spaces will not redirect measurements correctly.

* Substitute _ for spaces to run appropriately.

*

* 2. Duplicate image (Image > Duplicate...)

* make title "working_copy"

*

* ** Steps 3-7 are performed with the workingCopy image selected

*

* 3. Convert to 32-bit format (Image > Type > 32-bit)

*

* 4. Normalize intensity values (Process > Enhance Contrast...)

* set saturated pixels to 0.01%

* check normalize box

*

* 5. Subtract background (Process > Subtract Background...)

* check box for Light Background

* check box for Sliding Parabaloid

* adjust Rolling ball radius for best result (this macro uses 256)

* use preview check box to adjust rolling ball radius

*

* 6. Normalize intensity values (Process > Enhance Contrast...)

* set saturated pixels to 0.01%

* check normalize box

*

* 7. Segment by local minima (Process > Find Maxima...)

* check box for light background (changes from local maxima to local minima)

* select output type as Maxima within Tolerance

* adjust noise value to optimize particle identification (range is 0-1, this macro uses 0.2)

* * note segmentation will begin from the intensity value

* of the local minima +/- the noise value

*

* ** Remaining steps are performed on the binary image returned from step 6

*

* 8. Fill holes (Process > Binary > Fill Holes)

*

* 9. Set desired measurement parameters (Analyze > Set Measurements...)

* for intensity information, the measurements can be redirected to another image

* (this macro redirects intensity measures to the original image)

* select desired options

*

* 10. Analyze the particles (Analyze > Analyze Particles...)

* Customize which particles are analyzed by using size and/or circularity cuttoff ranges

* (this macro uses all sizes and all circularities)

* Select what output image to show if any (this macro uses ellipses)

* check display results

* check clear results

* check add to manager

*/

/* Start Macro version here

*

*/

// open original image

// grab title of original image for redirect of intensity measures

origTitle=getTitle();

// make duplicate to work from

run("Duplicate...", "title=working_copy");

// convert from 8-bit to 32-bit to facilitate normalization

run("32-bit");

// normalize intensity values to 0-1 range

run("Enhance Contrast...", "saturated=0.01 normalize");

// subtract background to enhance particle contrast

run("Subtract Background...", "rolling=256 light sliding");

// re-normalize intensities to 0-1 range

run("Enhance Contrast...", "saturated=0.01 normalize");

// find local minimas and segment within noise tolerance

// this generates new binary image with inverted

// LUT displaying black particles on white background

run("Find Maxima...", "noise=0.2 output=[Maxima Within Tolerance] light");

// make all particles solid by filling holes in binary image

run("Fill Holes");

// set desired parameters to be measure with analyzed particles

// all parameters are enabled here and redirected to original image for intensity values

// customize as necessary

run("Set Measurements...", "area mean standard modal min centroid center perimeter bounding fit shape feret's integrated median skewness kurtosis area_fraction stack display redirect="+origTitle+" decimal=9");

// Analyze the particles to get measurements and ROIs

// generates results table with all measured parameters in columns and particles in rows

// also generates ellipse outline image of particles.

run("Analyze Particles...", "show=Ellipses display clear add");

Also from the results you showed Peter, it looks like a watershed operation could greatly help you. Try taking the binary image and running Process > Binary > Watershed. This will break up your particles at the weakest connection points.

Hi, Kristopher,

Thanks a lot for the code. I applied both the code and the 'watershed' to the two sample pics. From what I got, I think 'watershed' can detect the size of particles better; and for my application, I don't just concern with the distribution, also want to have the edge-information as details as possible. Cause as you said watershed just break the particles at the weakest connections points based on the non-modified image.

And I wondered if there is any other method can even detect the inner circuit within particles with heterogeneous structure (just like I mentioned to Peter), not just confined to the solid particles, in this way, the particles will not need to be broken up.

I am just novice to ImageJ, even to image analysis, really impressed with the power of ImageJ, and highly appreciate your guys help.

Hi Jeannie,

Responding to your earlier question about possibly overestimating the size of particles, this is one of the things I find least intuitive in ImageJ….

Basically, ImageJ might be measuring the area of the particles only (just the black pixels, in your case), or it might be measuring the area of the black pixels + any fully-contained holes (white pixels). This will depend on exactly which settings you are using, so it is best to check until you are satisfied. Hopefully some of the following could help:

- I *think* that ImageJ’s ‘Analyze Particles’ generally measures what it should according to the ‘include holes’ option, as described here:

http://imagej.nih.gov/ij/docs/menus/analyze.html#ap

In other words, you can choose with this setting whether ImageJ measures the area of the particle + the area of the hole inside, or the area of the particle alone. *However* (and unfortunately) the ROI that it creates is always only the outer boundary. Therefore if you measure the ROI afterwards, the area of the hole will always be included in the result. Unless, that is…

- Under ‘Analyze -> Set Measurements…’ you can choose the ‘Limit to threshold’ option. This means that, when areas are measured, only above-threshold pixels should be counted. Consequently, if you measure areas on an image in which the threshold is set (i.e. you can see the red regions) then only the pixels shown red will be included. This lets you use the ROIs provided by Analyze Particles to measure areas without holes.

I suggest trying some different combinations of this, probably for measuring single ROIs, and seeing what results you get. Also, an extra trick that may help: ‘Analyze -> Histogram’ will give you a count of the black / while pixels within your image (press the ‘List’ button to access the numbers). Since areas are basically considered as counts of pixels* then you should find that the total number of black pixels in the histogram should correspond with the total of all the areas Analyze Particles gives in its results table. If the results table gives something higher, then it must be including some white pixels too.

Pete

*-Will, areas are pixel counts * the area of a single pixel… found under Image -> Properties… I’m not sure if you have the pixel size available for your images.

Something else you might want to consider is how a ‘particle’ is defined in the first place… which involves the idea of ‘connectivity’.

The main question is whether two pixels are considered part of the same particle if they are only diagonally adjacent to one another…. if not, they must be vertically or horizontally adjacent. I have attached 2 images to show what I mean - distinct particles are shown in different colours. If diagonal connections are allowed (‘8-connectivity’, because there are 8 neighbouring pixels) then the image contains only 1 particle. If not, the image contains 3 particles (‘4-connectivity’). Areas are also affected.

I think Analyze Particles… uses 8-connectivity, i.e. fewer particles with larger areas. If you want 4-connectivity instead, I’m afraid you will need to use another command.

One option is to install Fiji - www.fiji.sc - which is basically ImageJ with many extra features and plugins added. This provides a Plugins -> Process -> Find Connected Regions command, which allows you to decide whether diagonal connections count or not:

http://www.longair.net/edinburgh/imagej/find-connected-regions/

This is really intended for 3D data, but it works for 2D images as well. You might need to run Edit -> Invert first to switch your black and white pixels.

Unfortunately it does not output so many measurements, but it can output counts of pixels (i.e. areas, kind of) and a labelled image (i.e. each region has its own unique integer value). To visualise this image more nicely, try Image -> Lookup Tables -> glasbey.

If this alone is not enough to separate connected particles, then the watershed command can be really useful, as Kristopher suggests. Note that Process -> Binary -> Watershed has a preference for *round* particles, so if your particles are not necessarily round then this could cause too much splitting. In that case, with Fiji you can install the BioVoxxel Toolbox, which has a nice ’Watershed Irregular Structures’ command:

http://fiji.sc/BioVoxxel_Toolbox#Watershed_Irregular_Structures

Peter

Thanks for the tip on BioVoxxel Toolbox

Peter, thanks for the introducing of the Biovoxxel Toolbox, I found ImageJ is like a treasure box, lots of potential of image analysis can be achieved by using it.

For interested readers, some valuable information is available in relation with this issue by following this link: https://www.researchgate.net/post/how_to_get_grain_boundary_map_from_sem_images#58b83a783d7f4bb5931cc20b

Hi Jeannie Wang, 

How did you finally analyze the particle distribution in the picture you attached? I am facing a similar issue .

Thanks!