Darko Ninković

Medical microwave imaging (MMWI) is emerging as a promising noninvasive diagnostic modality, offering a safe, portable, and cost-effective alternative to conventional imaging methods. Despite its potential, MMWI poses significant challenges due to the ill-posed and nonlinear nature of the underlying inverse scattering problem. Two main strategies are typically adopted: qualitative imaging, which yields rapid but approximate reconstructions, and quantitative imaging, which aims to recover the full permittivity distribution of tissues. Both approaches substantially benefit from accurate patient-specific background models that enhance reliability and reduce computational complexity. Building upon preliminary results presented at EuCAP 2025, this work advances the development of patient-specific head models for quantitative brain imaging. The study employs anatomically conformal basis functions and extends the model to include cerebrospinal fluid, which represents a strong scatterer with markedly distinct dielectric properties. The proposed framework improves reconstruction accuracy and convergence efficiency, thereby contributing to the advancement of quantitative MMWI toward practical clinical applications.

Open hardware for electrostatic discharge pre-compliance immunity testing is presented. This system is based on a digitally controlled high-voltage source featuring a custom-designed step-up transformer, a capacitive voltage multiplier, and an innovative printed pulse shaper. The voltage multiplier consists of a cascade arrangement of capacitors and diodes, allowing it to provide a constant output voltage in the range from −15 kV to 15 kV for the pulse shaper in accordance with the IEC 61000-4-2:2025 standard. The high-voltage source is designed for burst outputs, capable of stepping up constant voltages from 5 V to ±15 kV. The novel pulse shaper, printed on FR4 substrate, was designed and optimized using circuit analysis and full-wave electromagnetic solvers to generate standard current pulses for contact electrostatic discharges. Measurements, verified in an accredited electromagnetic compatibility laboratory, indicate that the proposed open-hardware electrostatic-discharge generator produces the required voltages and contact-discharge current pulses within standard tolerances. The design choices prioritize ease of reproducibility, manual assembly, and the use of widely available components. All production files and accompanying microcontroller software are publicly accessible and released under the CERN Open Hardware License to foster innovation, promote collaborative design practices and enable broader access to electrostatic discharge pre-compliance immunity testing.

One of the main challenges in medical microwave imaging is to provide meaningful images when only minimal a priori information is available. Concerning the case of brain stroke imaging, this paper proposes a novel method for generating a low-resolution patient-specific approximation of the head by processing the same data used for the diagnosis. The method assumes knowledge of the head shape and a rough approximation of the fat and skin tissue above the skull. The approach consists of two steps. First, the measured data are processed using the first-order Born approximation scattering model to generate a qualitative head image. The underlying linear inverse problem is regularized by representing the unknown complex permittivity in terms of Chebyshev polynomials to improve its stability and effectiveness. Then, the voxels belonging to the skull are identified in the obtained image, and the skull boundaries are approximated from the discrete data using spherical harmonics. The proposed method is tested against an anthropomorphic phantom, demonstrating its capability to derive a patient-specific approximation. Moreover, the obtained background model enables stroke detection using real-time qualitative imaging.

Comparison of machine-learning antenna models, based on one deep neural network, an ensemble of neural networks and a mixture of experts, is presented. Two datasets comprising 100,000 samples each, with the sole difference in the ranges of antenna parameters, are used for training of considered machine learning approaches. Comparison in terms of accuracy and training time show a significant influence of the chosen design space (dataset) for training. One order of magnitude higher accuracy is achieved using a dataset generated with narrower ranges of antenna parameters. The newly proposed approach based on a mixture of experts achieved superior results when trained using wider ranges of parameters. If models are trained using narrower ranges of antenna parameters, then every model provides a unique compromise between the accuracy and training time.

We outline the ongoing research efforts in machine learning applied to electromagnetics at the School of Electrical Engineering, University of Belgrade, Serbia. Namely, in the past two years, our group has published multiple results regarding machine learning for antenna design and optimization, as well as applications for analyzing the emanated electromagnetic field from flat-panel monitors. We focus on finding the optimal topology of multilayer perceptron (neural) networks for the considered electromagnetic systems and assess the accuracy and efficiency of such models. The first reported example includes a dataset of up to 10 million Yagi-Uda antennas with four design parameters. The second example contains measurements of the emanated field from multiple flat-panel monitors over a time period of up to 48 hours. In both examples, we use ensembles of multilayer perceptron networks to overcome the modeling errors due to the stochastic nature of network training.

We present a method for identification of type and state ( on/off ) of flat-panel monitors, based on artificial neural networks and measurements of electromagnetic emanations in the presence of real-life noise, including other monitors in the vicinity. The proposed approach can identify monitor type and state with 99% accuracy (from a set of twelve). Instead of a single network, we use an ensemble of independently trained multilayer perceptron networks, since each training yields a different network. Further, we present an approach for automatic detection of significant frequency subranges of measured emanations, which eliminates less relevant inputs to neural networks and leads to reduction of their topological complexity, accelerates the training, minimizes the required set of experimental data and provides an overall insight into the spectral characteristics of emanations.