Various remarks (Japan course)

What I noticed about the NEUBIAS training resource

I’m Yusaku Ohta.

This resource, the NEUBIAS training resource, is really great, very informative.

I thought I knew these things very well, but when I actually tried them, there were a lot of things I didn't understand, and it's very useful. Especially, plugins like morpholibJ and clij are really useful and I will use them a lot for bioimage analysis from now on.

I have listed below what I was concerned about and what should be added. I will gladly edit these modules and create new modules if you allow me to do so. I am confident that I can make these modules even better.

Lookup tables It's also highly recommended to use the exact same color mapping for images that’re acquired on the same day and with the same microscope settings.

Multicolor lookup tables such as heat map, and fire, are also a good way to bring out small differences in the image, such as the FRET and FLIM image where images might have a narrower numerical range so that people understand the subtle differences across that range by multicolor LUTs.

It's important to avoid the red lookup table for those who are red-blind and use the magenta lookup table instead.

Data types 8 bit and 16 bit images do not perform calculations correctly if they involve negative values or decimal points in the answer. The big problem here is that all software does this calculation without warning us.

Image data types You write "unsigned N-bit integer images can represent values from 0 to 2^(N-1)". I think this formula is wrong and we don’t need parentheses in this formula. For example, if N equals 8, then 2 to the power of seven equals 128, but 2 to the power of eight equals 256. So we don’t need these parentheses.

Image file formats We have to note that converting file formats results in a loss of information. For example, saving a TIFF image as a JPEG changes the image dimensions, data type, pixel values, and metadata. And that means we can't get back our raw data here. Therefore, to re-save images for quantitative image analysis, we should choose a lossless compression file format like TIFF. It's a fairly old file format. But it supports both 8, 16 and 32 bit images and grayscale or RGB, which is nice because we work with them. It supports lossless compression, which is also nice because we don't want to use lossy compression.

Segmentation Image segmentation is the process of partitioning an image into objects or regions, which is required for the measurements of features such as intensity, area, volume and so on. So it’s really important in bioimage analysis.

Semantic segmentation divides images into regions of the same category, for example, Endoplasmic reticulum, mitochondria, cytoplasm and so on. It’s implemented in Fiji as a plugin such as Trainable weka segmentation. Instance segmentation ,on the other hand, divides images into individual objects, for example, “nucleus No.1”, “nucleus No.2”, “nucleus No.3”, and so on. It’s also implemented in Fiji as a plugin such as StarDist.

Auto Thresholding Many different algorithms basically work by assuming that our image is made up of background pixels and foreground pixels and that they are both roughly Gaussian distributed and they basically look for the minimum between those two distributions and then set the threshold there.

Global background correction Exercises No.2 I think line 19 of this code is wrong. run("Intensity Measurements 2D/3D", "input=intensity labels=labels mean max numberofvoxels"); -> run("Intensity Measurements 2D/3D", "input=xy_16bit__nuclei_with_background.tif labels=labels mean max");

Neighborhood filters The larger the filter size, the worse the spatial resolution of the image. For example, two nearby cell nuclei will be touching together. That's why we want to keep the filter size as small as possible.

NEUBIAS / training-resources

Various remarks (Japan course) #479