3D-image reconstruction of blood-flow detection experiments using neural networks

Laser speckle imaging (LSI) is a high-resolution, non-invasive, medical imaging technique that visualises motion of particles (e.g. red blood cells). The technique is very well suited to locate blockages of blood veins or asses the severeness of burn wounds. A big disadvantage of LSI is that it can’t produce 3D-images, like for example MRI- or CT-scans can. Here at the Physical chemistry and Soft Matter group we are developing a 3D-LSI set-up.

Figure 1
Figure 1

A central challenge in this project is the reconstruction of the 3D-images from a huge collection of different illumination and detection patterns. The reconstruction is a bit more complicated compared to the other medical imaging techniques because of the long path of the radiation bouncing through the sample. The X-rays in a CT-scan travel in a straight line which making the reconstruction straightforward. However in LSI the photons scatter through the sample so often that they follow a random-walk like path which is much less localized than a straight path.

Currently we can reconstruct 3D-LSI images by calculating all the possible light paths. However, this process is very slow, which makes it unsuitable for direct application in a hospital. A neural network approach will be able to speed up the reconstruction, broadening the scope and impact of LSI on the medical field massively.

Figure 2
Figure 2

Within this project, you will be designing this neural network and training it. To train the network we will need large amounts of data which we will generate using simulations. After training, we will apply your network to experimental data from our 3D-LSI setup. If you’re interested in a project that focusses on programming, but with both fundamental physics and applied medical imaging in mind, let me know!