Microwave Dielectric Property Reconstruction of Healthy and Tumor Rat Breast Tissues From In Vivo Measurements Using Generative AI

Abstract Objective. Rapid identification of suspicious breast lesions remains challenging, since pathological analysis following biopsy requires hours to days, limiting real-time decision-making, particularly during surgical procedures. Open-ended coaxial probe (OECP)-based methods have shown promise for characterizing target tissues through estimation of dielectric properties (DPs) from the measured reflection coefficient (S11). Conventional DP calculation from OECP measurements S11 is commonly performed using iterative methods, while recent studies have also explored artificial intelligence (AI)-based approaches for DP estimation. In these studies, AI models were developed using either synthetic datasets or limited ex vivo measurement data. In contrast, this study presents a generative AI approach for direct estimation of frequency-dependent DPs from in vivo measurements of healthy and tumor rat breast tissues. Approach. Measurements were acquired over the 0.5–6 GHz frequency range. Conditional Generative Adversarial Network (cGAN) and Conditional Variational Autoencoder (cVAE) were employed to learn the nonlinear mapping between the measured S11 responses and the corresponding relative permittivity (ε′) and dielectric loss factor (ε′′). Furthermore, model performance was evaluated using dependency-preserving metrics, including copula log-likelihood and Kendall’s τ , together with regression metrics such as root mean square error (RMSE), symmetric mean absolute percentage error (sMAPE), weighted mean absolute percentage error (wMAPE), and coefficient of determination (R2). In addition, the estimated DP responses were fitted to a one-pole Cole–Cole model using median normalized nonlinear least squares optimization to extract the corresponding parameters (εs, ε∞, τ , α, σ). Main results. The lowest RMSE values for healthy and tumor tissues were 0.98 and 2.90, achieved by the cVAE and cGAN models, respectively. The mean absolute percentage deviations, ∆(εc), between the Cole–Cole-fitted and reference DPs were 2.99% for healthy tissues and 5.05% for tumor tissues. Significance. These findings demonstrate the feasibility of generative AI-based, non-iterative DP reconstruction directly from in vivo OECP measurements.