基于径向基函数神经网络的脑部电阻抗断层成像仿真研究

Simulation study of brain electrical impedance tomography based on radial basis function neural network

目的:研究不同实现方式的径向基函数神经网络(radial basis function neural network,RBFNN)在具有真实解剖结构的脑部模型下的电阻抗断层成像(electrical impedance tomography,EIT)能力,为实际成像方法选择提供参考。方法:利用COMSOL Multiphysics软件基于脑部CT图像构建具有真实解剖结构的多层脑部EIT仿真模型并生成EIT仿真数据集,基于该数据集探究精确RBFNN、基于正交最小二乘法的RBFNN、基于K-Means的RBFNN 3种成像方法对图像重建结果的影响。采用图像相关系数(image correlation coefficient,ICC)和均方根误差(root mean square error,RMSE)评价成像结果。结果:无噪声的情况下3种方法均可成像且精确RBFNN的成像效果最好,在测试集中ICC和RMSE的平均值分别为0.784和0.467。基于正交最小二乘法的RBFNN在隐藏层节点数为50时成像效果最好,其ICC和RMSE的平均值分别为0.788和0.462。基于K-Means的RBFNN在噪声水平为30、40、50、60、70、80 dB时均能获得较好的成像结果,且能保持较稳定的ICC和RMSE,具备较好的鲁棒性。结论:3种RBFNN均可用于脑部EIT图像重建,且各有优劣,可根据实际情况选择合适的RBFNN用于EIT重建。

Objective To study the ability of radial basis function neural network(RBFNN)with different implementations for electrical impedance tomography(EIT)under real brain shapes,to evaluate the advantages and disadvantages of different approaches,and to provide a reference for the selection of practical imaging methods.Methods COMSOL Multiphysics was used to establish a multilayer 2D model with real structure based on brain CT and an EIT simulation dataset.The effects of the exact RBFNN,the orthogonal least squares-based RBFNN(OLS RBFNN)and the K-Means-based BRFNN(K-Means RBFNN)on the image reconstruction result were explored with the dataset constructed.The root mean square error(RMSE)and image correlation coefficient(ICC)were adopted to evaluate the imaging results.Results EIT could be completed with all the three RBFNNs without noise,and the exact RBFNN had the best results with average ICC and RMSE of 0.784 and 0.467,respectively,in the test set.The OLS RBFNN had the best imaging results at a hidden node of 50,with an average ICC and RMSE of 0.788 and 0.462,respectively.The K-Means RBFNN achieved the best imaging results at noise levels of 30,40,50,60,70 and 80 dB with stable ICC and RMSE and high robustness.Conclusion All the three RBFNNs can be used for brain EIT image reconstruction with their own advantages and disadvantages,and the RBFNN has to be selected for EIT reconstruc-tion based on considerations on actual conditions.