Abstract
Breast cancer remains one of the leading causes of death among women worldwide. One major challenge in early and accurate detection is breast density. High breast density not only obscures tumors on current imaging modalities, making them harder to identify, but also significantly increases the likelihood of diagnostic errors, both by medical professionals and automated detection systems. As a result, accurately classifying the breast density is crucial, and can lead to better, more tailored screening approaches and reduce the chances of error. This is especially critical for younger women, who are usually excluded from national screenings due to concerns such as radiation exposure. Microwave imaging offers a promising solution to this problem. Unlike traditional imaging methods, it uses safe, non-ionizing radiation, making it suitable for women of all ages. Beyond its safety, microwave imaging has the potential not only to detect breast cancer, but also to classify breasts into high or low density. This dual capability allows for more personalized and accurate cancer detection based on breast density, improving outcomes and reducing diagnostic uncertainty. Our microwave imaging prototype called MammoWave works by scanning the breast using a wide range of low-power electromagnetic signals captured from multiple positions around the breast. This approach provides a rich set of data that helps create an internal map of breast tissue without exposing patients to harmful radiation. This technique makes it possible to extract frequency-based characteristics from both the spatial and spectral domains, taking advantage of not just the signal's magnitude but also its phase information. These rich features can offer deeper insights into tissue composition and improve the accuracy of breast density classification. Our analysis shows that by fusing features from both the magnitude and phase of the signals-and focusing on approximately the first 40 components of the fast Fourier transform (FFT)-it's possible to achieve an accuracy of around 70% in classifying breast density using a support vector machine (SVM) with a radial basis function (RBF) kernel. Furthermore, instead of using the full frequency range (1 to 9 GHz), selecting specific sub-bands (1, 3, 4, 5, and 6 GHz) can improve the accuracy to approximately 73%. Importantly, the results also reveal that when breast density is correctly identified and taken into account, the performance of machine learning models in detecting breast cancer improves significantly boosting specificity and sensitivity by around 10% and 5%, respectively for low-density breasts, and by 15% and 10% respectively for high-density breasts.