Effect of frequency-dependent dielectric properties features on the accuracy of cancerous liver tissue identification.

Differences in the dielectric properties of normal and cancer tissues provide invaluable information for distinguishing them. The open-ended coaxial probe technique is a recognized method for measuring the dielectric properties of normal and cancerous tissues. Machine learning is widely used to classify these tissues on the basis of their dielectric properties. However, the current classification technique is error‑prone, as selecting an insufficient number of features may result in diminished accuracy, whereas selecting an excessive number can lead to overfitting in classification models. This study systematically investigated the impact of noise and individual variability, with particular emphasis on how the number of frequency points, which represents the features, influences the accuracy of cancer identification. The K-nearest neighbor (KNN) algorithm was used to analyze the dielectric properties, and the model was generalized to phantoms. Results demonstrate that as the number of frequency points increases, accuracy stabilizes and improves. However, accuracy gradually plateaued and slightly decreased with excessive frequency points, suggesting potential overfitting. A comprehensive analysis recommends using 20 frequency points for accurate tissue identification. When generalized to phantom dielectric properties data, the model misclassified only one phantom. The dielectric properties offer valuable insights into medical diagnostics. The KNN was effective in distinguishing between normal and cancer tissues. This study also provides practical recommendations for cost-effective dielectric properties analysis.