# Background Subtraction – Gaussian Filters

Performing background subtraction with a Gaussian filter can be an effective method of resolving
objects from a noisy background. More info here:
[https://bioimagebook.github.io/chapters/2-processing/4-filters/filters.html#gaussian-filters](https://bioimagebook.github.io/chapters/2-processing/4-filters/filters.html#gaussian-filters)

![Left to right: Image, Gaussian filter, Composite](images/gaussian-filter-example.png)

Background Subtraction - Gaussian filters
Performing background subtraction with a gaussian filter can be an effective method of resolving
objects from a noisy background. More info here:
https://bioimagebook.github.io/chapters/2-processing/4-filters/filters.html#gaussian-filters
1. Use **File > Open Samples > Neuron (5 channel)** to open the starting image. We only
want one channel to work with, so right click and **duplicate channel 4**.
- Try a variety of thresholds on this image, including the triangle method. Here we
will try to improve this threshold by separating out the smaller point-like
structures.
2. Open the gaussian subtraction script found here, developed by Ed Evans:

   [https://github.com/elevans/fiji-scripts/blob/main/imagej2/filters/ijo_gaussian_subtraction.py](https://github.com/elevans/fiji-scripts/blob/main/imagej2/filters/ijo_gaussian_subtraction.py)

(Download raw file in GitHub.)

Open this in Fiji to open the Macro Editor, or open the Macro Editor via **Plugins > New > Macro**

This script is written in python, so be sure to change the macro language to
python before running, otherwise an error will be produced:

![Changing macro language](images/macro-language.png)

3. With the single channel image highlighted, run the script. You should be prompted to
input a sigma value for the gaussian filter. Larger values will blur out larger and larger
objects. For this example, use 20 and then use 1. Feel free to try a variety of values to
see how the results change.

![Gaussian subtraction script interface](images/gaussian-sub-1.png)

4. After running the script, we now see that some structures are highlighted a bit differently
in result (V) (notice how the cell body is harder to see)

   - Some structures are highlighted differently.
   - The cell body may be less visible, while smaller axons or dots stand out more.

![Gaussian subtraction script output](images/gaussian-sub-2.png)


5. After running the script, we now see that some structures are highlighted a bit differently
in `result (V)` (notice how the cell body is harder to see):

![Gaussian subtraction thresholding params](images/gaussian-sub-3.png)
![Gaussian subtraction after thresholding](images/gaussian-sub-4.png)

6. If we decide this mask is sufficient for our data, we can then apply the mask to the output
image and merge the display.
a. With the output image still selected, choose **Apply** on the threshold window,
while making sure the dark background box is selected. It then will ask you to
convert the output to an 8-bit mask.
i. The areas of interest should now have values of 255 while the
background has a value of `0`. If this is inverted, it is likely that the dark
background box in the threshold window was not selected, you can simply
use **Edit > Invert** to switch the values.
b. To overlay the mask on the original image, we can make a composite. This
requires that the original image is also converted to 8-bit for display. Use **Image >
Type > 8-bit**

c. To make the composite, use **Image > Color > Merge Channels**. Here, I will set
the original channel duplicate to C4 (gray) and the mask to C2 (green). This
should procedure the following composite image, that can be used for display:

![Gaussian subtraction composite, Sigma=20](images/gaussian-sub-composite-20.png)

d. As an additional display, we could use **Edit > Selection > Create Selection** on the mask image,
then **Edit > Selection > Restore Selection** on the original image to get outlines of the mask.

If we use a sigma value of 1 and repeat the process, we can see there a further differences in
the cell body and other structures, although this seems to also pick up more noise:

![Gaussian subtraction composite, Sigma=1](images/gaussian-sub-composite-1.png)