RF-BreastNet: An Augmented Hierarchical Feature Extraction Framework for Breast Tumor Detection, Localization, and Size Classification From RF Microwave Signals

Radio Frequency (RF) microwave imaging is emerging as a modality for breast tumor diagnosis due to its non-invasive, non-ionizing, cost-effective, and wearable nature. Deep learning has shown great potential for imaging applications compared to traditional and complex image processing methods. But, deep learning algorithms require substantial amounts of labeled RF data for training, which are often limited in breast tumor diagnostic tasks such as detection, localization, and size classification. In the presence of limited labeled datasets, data augmentation becomes a crucial preprocessing strategy to enhance training effectiveness. We propose RF-BreastNet, an augmented hierarchical feature extraction framework for multi-task breast tumor detection, localization, and size classification in MRI-derived signals from female breast phantom. Our proposed framework effectively processed the dielectric properties of breast tissue with minimal error by leveraging the multimodal spatial, temporal, and metadata features of S11 and S21 parameters. In addition, we introduced a biologically informed augmentation strategy to improve generalization and reduce overfitting during training. The results indicate that RF-BreastNet framework demonstrated enhanced performance, with 97.19% accuracy in detection, 93.37% R² Score in localization, and 98.35% accuracy in size classification. The outcome of this study will facilitate early diagnosis and provide continuous breast tumor monitoring with high accuracy.