Bioimage Informatics 2023

Basic Information

The code is mainly used for BII2023 homeworks and group project. Please refer to BII2023 for more details.

This is an example Maven project implementing an ImageJ command with a Pom adapted for BioImageIformatics at EPFL. It was initially cloned from https://github.com/BIOP/ijp-template-ij2 (14th September 2022). The basic programming framework was kept.

New codes were included for resolving bioimaging tasks (i.e., multidimensional data pixel-wise operation, segmentation...) presented in lectures.

TIRF Protein Tracking (Mini-project)

Code avaiable here.

Total internal reflection fluorescence (TIRF) microscopy is an advance technique for quantifying fluorescence signal in a hundred-nanometer thin axial section. In the context of particle tracking, it provides high contrast images of how particles diffuse, interact, and traffic within the plane.

Together with Kuzey Aydin and Chang Liu, we built a particle tracking workflow for TIRF microscopy images. Specifically, I focused on the development of particle detection and trajectory linking, and my teammates worked on trajectory analysis.

Detection

We can see from the raw data that particles feature diverse moving patterns. In some samples, particles are trafficking for a long distance, while in others, particles burst locally within a short time interval. To generalize the workflow, I included both local max and binarization-based methods. If particles remain roundish and show a sharp intensity burst, the local max filter would be applied; whereas if particles elogate and show an ambiguous intensity profile, the thresholding binerization would be carried out. Two methods share common preprocessing steps (exponential correction, DoG, gaussian 3D, tophat).

Trajectory linking

I designed a cost function for forming correct linking that represents particle trajectories. In addition to the distance term and the intensity term, I added the orientation term $O_{t+s}$. The orientation term is the result of convolution by a time filter $(0,0,0,0,0,0,3,3,4,4,5)$ (inverted). It plays as an indicator of future, giving a clue of which direction the particle is moving towards. A lower $O_{t+s}$ implies a higher chance that the particle locates at the position after few frames. In addition, we implement linking not only between frame $t$ and $t+1$. Frame skipping is possible as $s=1,2,3,4,5$.

$$ c(x_t, x_{t+s}) = (1-\lambda-\gamma)\frac{d(x_t, x_t+s)}{d_{\text{max}}}+\lambda\frac{|f(x_t)-f(x_{t+s})|}{f_{\text{max}}}+\gamma\frac{O(x_{t+s})}{O_{\text{max}}} $$

Furthermore, I also set a upper limit for linkage formation. Due to the fact that it's impossible for particles to have a long displacement within a short time interval, this constraint discontinues trajectories correctly and avoids false connections.

Other Available Plugins (Assignments)

1. Bleaching Correction

Code avaiable here.

The fluorescence intensity of a life-imaging sample may decay exponentially as time goes because of photobleaching. The bleaching process can be represented as

$$I(t) = A*\exp{(t/\tau)}+C$$

Given an image stack, this plugin fits the parameters $A$, $\tau$ and $C$. It would perform a pixel-wise computation to make the mean of each frame in a comparable level to compensate such gradual drop. A certain level of noise can be observed in later time frames due to higher SNR.

2. Multiple Channels Quantification

Code avaiable here.

The plugin computes the gene expression (in form of fluorescence) of the cytoplasm at the very close periphery of every nucleus in a sequence of images. The input should includes two stacks, one for nucleus and the other for cytoplasm. The plugin first generates a mask of nucleus and then overlays ROIs onto the cytoplasm image. The measured values would be plotted into a time vs. intensity scatter plot.

3. Particle Tracking

Code avaiable here.

Spots in a sequence of images are tracked and trajectories are plotted. The workflow can be divided into two parts: spot detection and association. Detection is done with a DoG filter followed by local max filtering. Association is done by evaluating the cost function

$$ c(x_t, x_{t+1}) = (1-\lambda)\frac{d(x_t, x_t+1)}{d_{\text{max}}}+\lambda\frac{|f(x_t)-f(x_{t+1})|}{f_{\text{max}}} $$

where $d$ is the distance between two spots, $f$ is the intensity of the spot, and $\lambda$ is a parameter to balance the two terms. In addition, potential division of particles is considered and assigned with the same color during visualization.