A Preliminary Study on the Impact of Model Complexity in Classification in Breast Microwave Imaging

This study investigates the capability of microwave signals to classify benign and malignant breast tumours. The novelty of this study lies in the anatomically and dielectrically accurate breast tissue models - including tumours - used in the simulations, in multiple scenarios with different tissue types (fat, fibroglandular, skin, and muscle) and varying heterogeneity. Principal Components Analysis (PCA) was applied as a Feature Extraction Method (FEM), and Support Vector Machines (SVM) for the classification. Key metrics including accuracy, sensitivity, specificity, F 1 score, and Matthews Correlation Coefficient (MCC), assessed each antenna's diagnostic capability and the combined performance across all antennas. Results suggested that combining classifications from all antennas via majority voting improved accuracy, while tumour classification was unaffected by tissue heterogeneity. Increasing model complexity did not impact SVM performance except when muscle tissue was added, slightly reducing the diagnostic performance.