Multi-Objective Deep Learning for Microwave Tomography Based Lymphedema Detection

We propose a method to detect and localize lymphedema using tomographic microwave backscatter measurements of the upper limb. Inspired by object detection techniques in ground-penetrating radar (GPR) systems, we reformat the set of tomography scans into a B-Scan representation for analysis. Prior approaches relied on forming backprojection images to isolate returns from the lymphedema, but these imaging approaches are limited by the low resolution of the system. In contrast, here we employ a deep learning model that processes B-scan images directly, effectively canceling background signals from the layered limb tissues. The model predicts lymphedema intensity as a function of the azimuth angle, a representation that ensures robustness against variations in limb profiles and the placement of the limb within the sensor array. We incorporate a hybrid loss function for multi-objective deep learning to control the false alarm rate explicitly. To validate the proposed detection strategy, we generated a new simulated limb phantom dataset comprising 1,410 distinct profiles. Our results demonstrate that the hybrid loss function can be successfully used to improve lymphedema intensity estimation for negative cases and challenging, positive cases when the radius of the lymphedema is small. Furthermore, our method bypasses the computationally demanding backprojection imaging step, providing a computationally efficient solution for microwave-based lymphedema diagnosis.