A Novel Reconfiguration of Micro-Strip Patch Antenna with CSRR Metamaterial for Breast Cancer Detection

Abstract

Microwave resonant sensing and explainable machine learning can provide complementary evidence for breast-cancer decision support, but the two evidence streams are frequently conflated. This study proposes a dual-state microstrip patch antenna loaded with a complementary split-ring resonator (CSRR) and an independently reproducible artificial-intelligence branch. The antenna uses PIN-diode reconfiguration to alternate between nominal 2.45-GHz and 3.55-GHz interrogation states. An equivalent-circuit model predicts resonance shifts of 166.0 MHz and 293.1 MHz, respectively, for a malignant-like contrast load relative to air; these values are analytical phantom contrasts, not measured diagnostic thresholds. The executable classification study uses the public Kaggle/UCI Wisconsin Diagnostic Breast Cancer dataset (569 cases, 30 image-derived nuclear descriptors). The descriptors are arranged in their native 3 × 10 mean–standard-error–worst map. A leakage-aware, equal-score TriFusion combines balanced logistic regression, RBF support-vector machine, and random forest. Ten repetitions of five-fold stratified validation produced an area under the receiver-operating-characteristic curve of 0.9957, accuracy of 97.95% (95% CI: 97.83–98.06%), balanced accuracy of 97.68%, sensitivity of 96.65%, and F1 score of 97.22%. It achieved the highest mean accuracy, balanced accuracy, and F1 among the same-protocol models; logistic regression had marginally higher sensitivity and the lowest Brier score, KNN had the highest specificity and precision, and RBF-SVM had a statistically indistinguishable AUC. A convolutional surrogate reached 98.25% fixed-test accuracy and 99.12% class agreement with the TriFusion teacher. Feature-map Grad-CAM was supported by an activation-deletion test: masking the five most salient cells caused a larger target-logit to decrease than random masking (one-sided Wilcoxon p = 9.59e-21). The resulting architecture is an evidence-bounded proof of concept rather than an end-to-end antenna-to-diagnosis system; fabrication, full-wave simulation, tissue-phantom experiments, and co-registered RF data remain necessary