Transfer Deep Learning for Dielectric Profile Reconstruction in Microwave Medical Imaging

Abstract

Quantitative medical microwave imaging based on deep learning (DL) faces the overfitting problem due to limited training samples available in the clinic database. In this article, a U-Net-like DL model that can reconstruct the dielectric properties of brain tissue using time-domain signals is presented. A transfer learning approach is employed to alleviate the overfitting problem caused by limited training samples. In the proposed approach, the model is first trained with a dataset of random objects in a defined imaging domain and the corresponding time-domain signals. Subsequently, the pre-trained model is fine-tuned using simulation data from an unhealthy object. The final trained model can accurately reconstruct various tissues and abnormal lesions in an unhealthy object and avoid erroneous reconstruction of unexpected lesions in a healthy image of the object. The method is tested using a 16-antenna head imaging system operating across the band 0.5-2 GHz. The results confirm the superior performance of the method, in imaging both healthy and unhealthy brains, as measured using the root mean squared error, the correlation coefficient, the structural similarity index measure, and the peak signal-to-noise ratio. The presented method is a potential solution to mitigate the problem of erroneously predicting lesions in healthy tissues.