School of
Information Technology and Electrical Engineering

Speaker: Ahmed Al-Saffar
Seminar Date: Thu, 20/02/2020 - 10:00
Venue: 49-502; AEB Seminar Room
Host: Prof Amin Abbosh

Seminar Type:  PhD Confirmation Seminar

Abstract: 

Electromagnetic imaging (EMI) is a technology attracting increasing interest in many fields. In medical applications, it has the benefits of portability, safety and affordability compared to conventional imaging modalities. At the core of this technology lies the inverse scattering problem (ISP) which seeks to reconstruct the dielectric properties of the imaging domain from measurements. Offering a stable and noise-robust solution to an ISP is an ongoing challenge. High dimensionality of the imaging domain, small number of measurements, noise in the environment and scatterings, make ISPs under-determined, ill-posed and highly non-linear.

Far from conventional iterative solutions, researchers have recently attempted to bring deep learning techniques to EMI. The speed and expressive power of deep neural nets (DNN) make them a good candidate for nonlinear ISP. That said, deep learning comes with its own set of unique challenges that seriously limits its applicability. The overarching concerns are: scarcity of data required for training, and the sensitivity of electromagnetic hardware to noise and variations in environment. This research aims to tackle those challenges and provide a functioning and robust algorithm in real life settings. While collecting real data at scale to train a DNN is unfeasible, a fair amount of data can be feasibly generated via computer simulations, however, it doesn't perfectly represent real environments and therefore, any neural net trained on simulation data is doomed to fail when evaluated on real data. In our work, we customized domain adaptation technique to match distributions of complex-valued electromagnetic data. An operational NN trained on simulation data and adapted to real data to perform target localization is presented. In addition to localization, we aim to tackle the overarching task of reconstruction of dielectric profile of human head. Initial results based on coarse 2D simulations are presented. Aiming at a functioning algorithm in real settings, we realize that pure learning approaches are too fragile to work in such a sensitive application, and thus we seek to combine the learning approach with the well-established physical-analytical approach of solving this inverse problem. Lastly, in medical diagnosis, life-threatening decisions are made based on the outcome. As such, we envisage a model capable of estimating uncertainty in the results. 

Biography: 

Ahmed Al-Saffar is a PhD student in the EMAGIN group in the School of Information Technology and Electrical Engineering at the University of Queensland. His supervisors are Prof Amin Abbosh, Dr Alina Bialkowski, Dr Lei Guo and Dr Mahsa Baktashmotlagh. Ahmed received his BSc (Hons) in Electrical Engineering from the University of Mosul in 2013. His research interests revolve around machine learning and signal processing. His research work focuses on applying machine learning techniques for electromagnetic imaging tasks in real-world settings.