Tomographic Ultrasound Imaging via Deep Learning for Breast Cancer Localization

In the framework of medical breast imaging, several technologies have been exploited throughout the last decades to provide qualitative and quantitative information about the inner tissues of the breast to support physicians in their investigations. Among these, ultrasound tomography stands out as a promising solution, combining the cost-effectiveness and safety of ultrasound waves with good spatial resolution typical of clinical tomography, such as computed tomography. This technology aims to generate maps of the acoustic parameters of internal breast tissues by analyzing pressure waves acquired by sensors placed around the patient’s breast. From a mathematical perspective, this imaging task poses a highly non-linear and ill-posed inverse problem, making its solution inherently challenging. Over the years, several approaches have been proposed, leveraging prior information and specific simplifications to address these difficulties. In this paper, we introduce a deep learning-based approach to tackle this problem. Specifically, an convolutional neural network is designed to process the scattering matrix of the acquired pressure waves and estimate the internal structure distribution of breast tissues rather than retrieving their mechanical features quantitatively. To evaluate the proposed method, a realistic, tailor-made dataset was generated and used for training the network. Initial results show the potential of the proposed deep learning approach, highlighting its potential impact in breast cancer imaging and its diagnostic utility.