Abstract

Info

Title:: Microwave imaging method based on deep convolutional autoencoder and its potential for medical application

Author(s):: DENG Huangsen; LIU Jie; YAN Lian; ZHU Guangzheng; NING Xu; QIN Mingxin; CHEN Mingsheng; School of Biomedical Engineering and Medical Imaging, Army Medical University, Chongqing 400038, China

Keywords:: Keywords: stroke microwave imaging deep learning convolutional autoencoder

Abstract:: A deep learning based microwave imaging model which can directly map the scattered electric field to the dielectric property distribution image of the target object is developed, and its potential for medical applications is explored. The two-dimensional time-domain finite difference method is used for numerical simulation to obtain a dataset of scattered electric fields a deep convolutional autoencoder based imaging model is constructed to perform imaging studies on two types of target objects the imaging results are quantitatively evaluated using relative error, and the models ability to distinguish different types of strokes is also analyzed. The results show that the imaging network based on deep convolutional autoencoder exhibits excellent imaging performance when processing both numerical models. For simple objects, the model can accurately locate and preliminarily reconstruct the shape of the object, with an average relative error of 0.301 2, while for the stroke models, it can effectively reconstruct the location and shape of the stroke area, and preliminarily reconstruct other brain tissues, with an average relative error of 0.077 8. The microwave imaging network based on deep convolutional autoencoder has great promise for fast and accurate image reconstruction, and the numerical example of stroke detection demonstrates its significant application potential in biomedical imaging.