Enhanced Learning in Microwave Medical Imaging using Boundary-Overlap-Size Loss

Loss functions are essential for optimizing the imaging performance of trained deep learning models. Thus, the design of tailored loss functions defines the effectiveness of deep learning imaging models. A loss function based on the three geometric measures of the local suspected lesion area and the global image quality of the reconstructed image is proposed. The function is used to efficiently train customized U-Net++ for reconstructing human brains’ dielectric profiles using microwave imaging modality. The input data for training is the broadband time-domain signals irradiated within an imaging domain by 16 antennas. The achieved results confirm accuracy of the trained model using the proposed loss function in reconstructing tissues images quality and lesion area identification.