1. Choosing the right parameters
Last updated
Was this helpful?
Last updated
Was this helpful?
It is recommended to test the GAT segmentation on representative images of neurons first before running the entire workflow. These macros can be found under GAT->Tools and will have "Test" as a prefix. This is to ensure that the segmentation is accurate.
There are 3 important parameters in GAT that may require fine tuning:
Probability
Rescaling Factor (dependent on pixel size)
Overlap Threshold
We will go into more detail about each and how to optimise them.
The general workflow would be to test different rescaling factors first, followed by different probability thresholds. There are scripts within GAT -> Tools that make it easy to test a range of values.
The models within GAT are trained to segment neurons with mean area of: 701.2 ± 195.9 pixel2 (Mean ± SD, 6267 cells). As images can be acquired at different resolutions (pixel sizes), it is important that the image be rescaled so the cells are of a similar area (in pixels) to what the model/s have been trained on. This is done by specifying a rescaling factor. By default, the rescaling factor within GAT is 1. However, you can test a range of values to ensure you use the most accurate value. If the rescaling factor is too big, it will split your cells, if it's too small, cells are merged. This is illustrated in the image below:
yellow arrowheads: merges
white arrowheads: splits
asterisk: false positive
Adult Mouse: 0.9 to 1.2
Adult Human: 0.5 to 0.7
If your cells are being over-segmented or split, increase the rescaling factor, if they are being merged, decrease the rescaling factor.
Please use a value that works accurately on your dataset.
To make testing easier, if you have a large image, you can crop it and test a small region.
Ensure your images are calibrated. Open your images in FIJI and go to Image -> Properties on the FIJI menu.
It should state the pixel width and height in microns. If not, please enter this manually. GAT will only work properly if this information is entered.
Go to GAT -> Tools -> Test neuron rescaling
Channel 3 is “Hu” labelling of the neuronal soma.
Click on Browse, navigate to the image and click OK.
Tip: If you already have an image open, you can tick the box Image_already_open.
As a start, its recommended to test a rescaling factor of 1. Enter the same value for minimum and maximum value to test only one rescaling factor.
Keep the rest as shown in the image above.
Click OK and it will open the image. As it’s a multichannel image, it will ask you to verify the channels so you can choose your channel of interest in the subsequent prompt.
Moving the slider on the bottom will cycle through the 3 channels. Hu is in channel 3.
Enter 3 in the next box and click OK.
It will now perform segmentation of the neurons using the settings entered above. As we are only testing one rescaling factor, we will get one resulting image with ROIs overlaid on it.
Keep in mind that this image has been resized and segmentation performed on this image to obtain the ROIs. Use this step to zoom into the image and verify the segmentation is working. If you’re not happy, you can test a range of values. If you’d like to test a range of values, close all the images and rerun “Test Segmentation”. In the dialog box, enter the minimum and max values. Here I’m testing 0.7, 0.8, 0.9, 1, 1.1, 1.2, 1.3 and 1.4. For this, I’ve entered
Minimum value: 0.7
Maximum value: 1.4
Increment step/s: 0.1 Click OK and enter channel 3 as above in the next prompt.
This will yield a series of images with different pixel sizes and the resulting segmentation overlaid.
The rescaling factor used will be in the image name. You can check the Log window to see the number of objects for each rescaling factor. It is recommended to test this on images you may have already counted or small regions manually analysed to verify the accuracy.
Default value in GAT is 1
Right click outside the ROI, click duplicate followed by OK and you will get just that region as a separate image.
Go to GAT -> Tools -> Test Neuron rescaling.
As the image is already open, tick Image_already_open.
Use the following settings:
Click OK. It will ask you to select the image as we selected the image already open box. Select the cropped image and click OK.
It will now cycle through the rescaling factors, resize the images and run segmentation on them.
In the Log window, you will see the no of detected objects at each rescaling factor with corresponding pixel size:
Earlier versions of GAT used pixel size for segmentation. This was confusing and instead we have introduced a rescaling factor. To convert from pixel size to rescaling factor you can divide pixel size by 0.568. Example conversions:
0.9
0.63
0.568
1
0.5
0.88
0.45
0.79
Keep in mind this is the pixel size you would like to rescale the image to so that the algorithm performs accurate segmentation. It is not the pixel size of your image.
For earlier versions of GAT, click to see conversions from pixel size to rescaling factor.
For this example, one of the will be:
You can test this by downloading this .
Tip: You can use for analyzing large 2D tilescan images as well
The above parameters are all entered into the StarDist plugin in Fiji. For more details about StarDist and how it works, refer to this .
Open a large tilescan image in FIJI. You can select Tilescan_GAT_ms_distal_colon_MP_hu.tif from the Select the rectangle tool and draw a rectangle around region you’d like to test the segmentation.
