AI for speckle sensing

Speckle-based non-invasive assessment of cardiac vital signs offers a promising avenue for inclusive health monitoring. This PhD research in imec’s Photonics for Health program proposes an interdisciplinary approach combining deep learning with signal processing and other machine learning techniques to extract and interpret cardiac signals from laser speckle patterns.

Laser speckle imaging captures dynamic interference patterns generated by coherent light interacting with biological tissue. The speckle videos are determined by absorption and scattering properties of tissue (light tissue interactions), pressure, flow and volume information (physiological interactions of speckle) and tissue movement as well as body movement artifacts (motion artifact interactions). The project aims at disentangling these, using AI methods to find biomarkers relevant to cardiovascular health, among others by looking into reverse engineering methods such as saliency maps. The results will contribute to improved estimation of blood pressure and oxygenation among others. We propose to explore methods such as convolutional neural networks, autoencoders and diffusion models, based on the acquired images, and possibly also on the derived signals.

The proposed work aims to develop robust algorithms capable of isolating cardiac-related features from noisy speckle data. By integrating domain knowledge from biomedical optics and cardiovascular physiology, the research will explore hybrid AI models for improved accuracy, generalizability and interpretability.

We offer you a challenging, stimulating and pleasant research environment, where you can contribute to international research on applying artificial intelligence to a challenging health context. This research will be a collaboration between the Holst Centre in Eindhoven, and the AI & Data department at imec Leuven.
Our ideal candidate for this position has the following skills:

• You have a master's degree in mathematics, computer science, biomedical engineering or related.

• Programming experience in Python

• Knowledge in signal processing and machine learning

• Understanding of cardiac physiology is considered a plus

• You have strong analytical skills to interpret the obtained research results.

• You are a responsible, communicative and flexible person.

• You are a team player.

• Your English is fluent, both in speaking and writing

Background information:

We propose a day of meetings and collaboration at the Holst every two weeks, carpooling is a possibility. The proposed research potentially has an impact across the Photonics for health program (blood oxygenation, blood pressure, heart rate variability...):

• using deep learning to link input and output directly, possibly in metalearning with other machine learning methods to achieve more accurate and robust outcomes

• using deep learning to understand the underlying problem and find new white box biomarkers e.g.

  • e.g. reverse engineering based on saliency maps in 1D or 2D CNNs to discover new features/biomarkers

  • diffusion models for denoising

  • investigating the latent space of autoencoders

There are two European funded projects currently running that will continue for several more years the PhD will have an impact on. These use both non-contact and contact SPG with multi-wavelengths and between different environments, i.e. hospital and automotive:

• STIMULUS: contact-SPG multi-wavelength device aiming at microvasculature monitoring. The use of AI in this case could enhance the learning from the collected signals providing additional information on micro vs. macrovasculature in a diseased population. AI also could be a good fit here since data collections will be performed in different hospitals, providing datasets with significant sizes.

• REALCARE:  non-contact SPG device for biomarkers monitoring in patients. Also, in this case we foresee a considerable dataset to be obtained with videos of patients and SPGs. AI could be used to automatically extract the SPG from the parts of the images that contain the best quality possible, learning from the images directly where the SPG that gives optimal results on heartrate can be found or where the SPG that contain the most information on respiration can be found. This will in turn allow the use of DL models to automatically provide from the videos the HR and RR needed as output in the project.