Goal of the Hackathon

The goal of this hackathon, besides having a good time (of course), is to learn something about the analysis of natural data in a practical application. You will be provided with sensor recordings and a code framework that shows how an analysis pipeline for detecting nanoparticles in liquid samples can be accomplished. You can follow the existing path, trying to implement the most effective or most efficient models in the given pipeline, or creatively develop a completely new approach. Some challenges such as low availability and imperfect training data, which regularly occur in real applications and imaging artifacts, will require additional thoughts on possible problems and solutions. In addition to that, for the given system, resource constraints also apply and must be considered explicitly in one of the tasks.

Motivation and Data

Detecting nanometer-sized particles like viruses in samples is an opportunity to quickly identify the spread of infections and thus enable measures to quickly counteract. In this scenario, we rely on the PAMONO sensor as a basis for the visualization of nanoparticles in liquid samples.

Although the images taken with this sensor contain the information necessary for particle detection, it is not sufficient to analyse each image on its own since the background signal of the images is very strong and particle attachments only reveal themselves through a temporal change of particle intensities.

After preprocessing a window of consecutive images, particles become spatially visible (marked red in the image):

And can be distinguished from images containing no particles:

The search for the characteristic shapes of particles is hampered by naturally occurring disturbances and their characteristics, which could change with each recording, as well as by the availability of only a small amount of training data. For this problem, an appropriate solution is required that is as robust as possible and, ideally, also resource-efficient to enable the use as a rapid test in a natural environment.

Visualization of the data recording process (click on the image to see the video):

Further explanations of the data characteristics:

• Presentation (PDF) of the data characteristics

Helpful information for all tasks

There are several things to keep in mind in all tasks.

• Beware of overfitting. There is relatively little training data, as is unfortunately often the case in medical applications.
• As with all sensor applications, disturbances will occur. See natural imaging data and artifacts for more information on this topic.
• You should respect and follow the syntax specifications of the individual tasks.
• You do not necessarily have to use machine learning approaches. You could replace those parts with different types of methods if they seem more promising.

Before you start

Have a look at the Libraries section for information on the allowed libraries.