Attention U-Net for Binary Mask Generation in Medical Microwave Imaging

Conventional microwave imaging as an inverse problem is ill-posed, non-linear, and requires large computational resources. To reduce that ill-posedness, an Attention U-Net model is proposed to generate a binary mask of the dielectric properties of the domain. The generated binary mask can be used to extract tissue regions, enabling the post-processing algorithms to focus on the target. In this work, the training images along with the ground truth labels are generated after a pre-processing of substantial simulation data. The ground truth labels are binarized before training. The testing results of the neural network indicate more than 82%, 70%, and 74% accuracy in the pixel properties estimation, Dice coefficient, and the mean intersection of the union. Those results indicate the potential of the proposed approach in medical microwave imaging.