Bacteria Density Analyzer: Abstract

0. User provided data

At first, the user is expected to provide several channels belonging to a 3D image.
The mandatory channel is the one we will use to segment the filaments. During the tests and the development of this module, we used a nuclei staining (DAPI).
The other channels are fluorescence channels corresponding to bacteria stains (for example GFP, RFP, etc.).
The module can handle an arbitrary number of fluorescence channels at the same time.
The image can contain several filaments, even if they are broken or partially segmented.

As images contain several filaments, the user must indicate what part of the image belongs to which filament.
To do so, the user must draw polygons around each filament using Napari’s Shapes layer.
The color associated to each polygon will define the filament’s ID.
A filament can be broken into several parts, each part being surrounded by a polygon of the same color.
On the example image below, there are two filaments/branches (the red and the green ones), the first being broken into three parts and the second being monolithic.
On the example, the three regions of the red filament are adjacent, but in reality, breaking a filament should only be done if a part of it can’t be segmented (for example because of low signal-to-noise ratio), so the different parts can be far from each other.
The plugin will by itself re-link the different chunks in the final output.

As a filament can be broken in any number of parts, the user must provide for each branch (== each color) one single point that approximately corresponds to the start of the filament.
From there, the plugin will rebuild the full filament by connecting all the different parts together.
The starting points have to be placed inside the corresponding polygons.
The exact location doesn’t matter as it is just a proximity indicator. The path of every filament will be processed and then, we will search for the closest point to the starting point provided by the user.

Everything starts by segmenting the filaments in 3D using the segmentation channel provided by the user.
A binary mask is created and fixed, allowing us to know which pixels are part of the filaments and which are not.
If a pixel is positive in the mask, its intensity will be used in the computation of the integrated intensity and density.

From the 3D binary mask, we compute a skeleton of the filaments.
It allows to reduce the 3D structure of the filaments to a 1D path going through the center of the filaments.
A skeleton is typically represented as a binary mask containing 1-pixel wide lines going through the center of the filaments.
We will base ourselves on this skeleton to compute the length of the filaments and to extract the intensity values along them.

Depending on the mask and its level of details, the skeleton can contain small branches and loops that are not part of the real filaments.
To avoid these artifacts, we extract a medial path going through the skeleton from one end to the other.
To extract this path, we start by extracting the pixel corresponding to the starting point for each chunk.
From this pixel that we consider as the root of the filament, we search for the longest path going through the skeleton.
This way, we avoid small branches and loops that could be present in the skeleton.

We now have a 3D mask indicating which pixels belong to filaments, and a 1D path going through the center of each filament.
The goal is to project (in some way) the 3D information onto a 1D object.
To do so, we take each 3D coordinate from the mask and search the closest point on the 1D path.
For each point of the 1D path, we create an accumulator that will store the integrated intensity (== the sum of all the intensities of the voxels that project onto this point).
Using the same principle, we can extract the volume of each accumulator by counting the number of voxels that project onto each point of the 1D path.
Finally, the density is computed by dividing the integrated intensity by the volume for each point of the 1D path.
Exporting such values would result in rather noisy curves, so we bin the values according to a user-defined length (for example 1 micron).
It basically means that we group the 1D points by chunks of 1 micron and gather the statistics by chunk.