Accelerated Medical Microwave Tomography via Heterogeneous Computing

In practical applications, such as medical microwave imaging, the implementation of three-dimensional (3D) numerical electromagnetic (EM) inverse scattering algorithms has been limited for over two decades owing to their high computational cost. The emergence of graphics processing unit (GPU) computing has offered new possibilities to alleviate the computational burden related to EM calculations; the optimized deployment of EM inverse scattering algorithms on GPUs or other accelerators, however, has not been explored. To this end, we present the first complete GPU implementation of an iterative EM inverse scattering algorithm, which employs a suite of optimization tools in both the forward and inverse problems solved at each iteration. Importantly, we study the impact of various implementation alternatives on computational savings in different computing platforms. Our results demonstrate the significance and benefits of jointly optimizing the forward and inverse problems in implementations in heterogeneous systems. Importantly, our optimization achieves the fastest fully numerical implementation of nonlinear, iterative, EM inverse scattering algorithms that has ever been reported in the literature, resulting in real-time, 3D MWT-based reconstructions. This can pave the way for the development of clinical systems for microwave-based applications such as breast cancer or stroke detection, which are designed to take full advantage of 3D MWT scanning and expand their imaging capabilities.