Improving the Diagnostic Capability of Microwave Radar Imaging Systems using Machine Learning

Breast microwave sensing (BMS) is a potential breast cancer detection technique that uses low-power microwave radiation to detect the presence of cancerous lesions. This work presents the results of the application of a multilayer perceptron (MLP) and support vector machine with radial basis function (SVM RBF) to breast cancer detection for a portable BMS prototype. Numerical 2D phantoms belonging to either BI-RADS Class 1 or Class 2 classifications were used to produce simulated data as collected by the portable system using an array of twelve sensors operating at five frequencies between 2.3 GHz and 6.5 GHz. Five feature preprocessing pipelines and their impact on classification performance were evaluated. An area under the curve of the receiver operating curve (ROC AUC) as high as (95 ± 1)% for BI-RADS Class 1 and as high as (94 ± 1)% for BI-RADS Class 2 were obtained using the SVM RBF, and as high as (94 ± 1)% for Class 1 and (92 ± 2)% for Class 2 using the MLP.