PhD - Leuven | More than two weeks ago
Terahertz (THz) tomography is an up and coming technology that uses electromagnetic radiation with terahertz frequency for tomographic imaging. Like X-rays, THz waves provide information about the interior of an object through interaction with the object. THz waves interact with many materials in different ways. They are absorbed in polar materials such as water, penetrate most packing materials (plastic, paper, ceramics, …) and are completely reflected by metal. In contrast to X-rays, there are no known negative effects of THz waves, making their application attractive for biomedical purposes as well as industrial inspection, non-destructive testing, material science and agro-food applications. The Gaussian THz beam however, diverges much faster than an X-ray beam and reflection and refraction effects play a dominant role, preventing the use of conventional X-ray reconstruction techniques.
Imec developed, on the one hand, THz emitters and detectors in their integrated photonics platform addressing the 0.1 to 2.2 THz range, and on the other hand, prior-knowledge based iterative reconstruction techniques for THz tomographic data. These techniques model the physics of the THz image formation, such as the THz beam shape, in the image reconstruction process, as opposed to performing pre- or post-processing steps. Such algorithms are still nearly unexplored for THz imaging and have so far increased the applicability of the technique through a substantial improvement in image quality. However, THz tomography remains a relatively slow imaging technique, plagued by noise and undersampling artefacts in the reconstructed images.
The use of deep learning to enhance tomographic reconstruction is rapidly emerging in the fields of material science, (bio-)medical imaging and industrial inspection. It often concerns image quality enhancement, which improves the fidelity of visual analysis of feature extraction by reducing image noise, blur, scatter, etc. While these often outperform numerical optimization methods in specific applications, they lack data consistency terms that link back to the measured data, which is crucial to reduce artifacts and hallucinated regions. We aim to address this limitation by development of a physics-informed deep learning technique, that is able to provide high quality reconstructions from few view data at high speed. Model-based deep learning methods form a new class of deep learning techniques which combines the data-driven approach of deep learning with a mathematical forward model which represents the underlying physics and typically require less training data than traditional deep learning techniques. The selected forward model needs to be differentiable to be integrated directly into a deep learning end-to-end reconstruction framework.
We look for a strong candidate to join our teams and develop a demonstration platform for THz spectral imaging. To that purpose, we offer access to our state-of-the art integrated photonics and nano-optics platforms as well as imaging and deep learning algorithmic know how.
Required background: Physics, nano-engineering
Type of work: 60% modeling / simulation, 30% experimental / data analysis, 10% literature
Supervisor: Jan Genoe
Co-supervisor: Xavier Rottenberg
Daily advisor: Roelof Jansen, Jan Sijbers
The reference code for this position is 2024-143. Mention this reference code on your application form.