91ɬÂþ

Abstract

Skin cancer is the fastest-growing form of cancer, accounting for one-third of all cases worldwide and representing a major public health threat. Visual assessment and invasive biopsy remain the clinical standard, but they depend on subjective perception and experience of the dermatologist, patient discomfort, and lengthy waiting time for diagnostic results.

Microwave techniques show promise as a non-invasive alternative by exploiting dielectric contrast between healthy skin and anomalous lesions. So far, the reported sensors based on this approach have been primarily reflectometry-based dielectric probes and antennas. These appeared to struggle to provide, within a single platform, deployment flexibility, cost-effectiveness, and quantitative dielectric characterization. In contrast, transmission- based approaches, particularly those using surface waves, show promising potential but remain largely unexplored.

The work of this thesis targets exploration of a surface-wave-based method to address these limitations in microwave-assisted skin cancer diagnosis. In this approach, two bio-compatible patch antennas are placed on opposite sides of the suspicious skin region. One antenna excites and the other receives surface waves propagating along the skin–air interface. The challenge lies in strategic analysis of the transmitted response to uncover the information about the dielectric profile, thereby detecting the skin anomaly.

Building on this hardware concept, a comprehensive, end-to-end frame-work is established. First, a theoretical model describes how surface waves interact with skin and lesions and links antenna phase shifts to tissue permittivity. Subsequently, the antenna evolution is shown, including one monopole-based design (Gen 1) and two Vivaldi-based designs (Gen 2 and Gen 3). Next, numerical simulation and phantom-based experiments are presented to validate the theory and compare the performance of several antenna designs. The results confirm that the surface-wave method can characterize permittivity contrast, detect the presence of skin lesions, and track their progression. A steady improvement in performance is observed across successive antenna generations. Finally, a deep-learning classifier for skin-condition prediction is trained on measurements from 900 tissue models that reflect person-to-person variability. Using a convolutional neural network (CNN) enhanced by a bidirectional long short-term memory (Bi-LSTM) block, the model achieves over 90% accuracy in flagging malignant tumors, demonstrating the potential of the surface-wave approach as a flexible, low-cost solution for skin cancer detection and monitoring.

In addition to the above-describe chief exploration, two important issues in microwave-based skin diagnostics are investigated. First, the safety analysis is conducted via numerical tools, to show that peak specific absorption rate (SAR) can vary by more than 80% across different body sites, highlighting the need to account for tissue variability and providing a basis for a comprehensive evaluation framework. Second, the work explores the trade-off between the computational cost and complexity of tissue models in numerical simulations, using the results to provide a recommendation for optimized balance between computational economy and meaningfully accurate simulation data.