Deep learning algorithms for brain disease detection with magnetic induction tomography

PURPOSE: In order to improve the reconstruction accuracy of magnetic induction tomography (MIT) and achieve fast imaging especially in the detection of cerebral hemorrhage, artificial intelligence algorithms are proposed to improve the accuracy of MIT inverse problem. METHODS: According to the standard geometric data of human head, a three-dimensional (3D) head model containing four layer tissues is established for brain image reconstruction of MIT. Four deep learning (DL) networks, including restricted Boltzmann machine (RBM), deep belief network (DBN), stacked autoencoder (SAE), and denoising autoencoder (DAE), are used to solve the nonlinear reconstruction problem of MIT, and the reconstruction results of DL networks and back-projection algorithm are compared. Finally, in order to verify the practical value of DL algorithms, the phantom experiment is carried out with MIT detection system. RESULTS: Using the nonlinear data learning ability of DL algorithms, the rapid and high-precision imaging of cerebral hemorrhage can be realized. Compared with the back-projection algorithm, the DL improves the artifact and the accuracy of the reconstruction image. The location and volume of bleeding can be reconstructed and the prediction time reaches 20 ms. Moreover, the anti-noise performance of the networks can reach 20 dB. CONCLUSIONS: The DL can effectively improve the reconstruction accuracy and prediction speed of the image when it is applied to the reconstruction of cerebral hemorrhage in MIT. This feasibility study MIT to be a potential technology for brain diseases to fully meet the needs of accurate, rapid, and low-cost clinical diagnosis and continuous monitoring.