3D CNN for breast density classification in microwave imaging

Breast density assessment plays a crucial role in mammography interpretation and breast cancer risk evaluation. Traditionally, this process relies on subjective radiological assessment, which can lead to variability in results. Recently, deep learning algorithms have shown great promise in automating breast density classification based on mammography images. In this study, we propose a deep learning-based approach for breast density classification using microwave imaging technology. We developed and trained a custom convolutional neural network using raw measurement data from a microwave breast test device to classify patients’ breasts as dense or non-dense, with radiologist assessments based on the BI-RADS classification system serving as reference. The dataset comprised 6709 measurements: earlier data (first 5920 measurements) were split into training (75 %) and validation (25 %) sets, while the most recent data (last 789 measurements) formed an independent test set. The measurements were collected using different devices from multiple sites. The model architecture was selected on the basis of a device-stratified Monte Carlo cross-validation, ensuring that data from different acquisition devices were proportionally represented in each fold. The selected model achieved a total accuracy of 80.1 % ± 0.7 % on the test set, with 79.4 % ± 1.8 % and 80.8 % ± 2.4 % accuracy for low- and high-density cases, respectively. Furthermore, bilateral consistency was identified as a key indicator of classification reliability, reaching 84 % of performance accuracy when verified. This work demonstrates that it is possible to aid clinical decision-making in screening programs with quantitative results based on non-ionizing radiation. • First application of 3D CNN to raw microwave data for breast density classification. • Validates linear correlation between breast tissue electric properties and density. • Leverages a large multi-site dataset of 6709 microwave scans. • Achieved 80.1% accuracy in binary classification on test dataset. • Enables clinical screening personalization and support via non-ionizing radiation.