An Attention-Based Deep Convolutional Neural Network for Brain Imaging With Electrical Impedance Tomography

Electrical impedance tomography (EIT) is a promising technique for brain imaging. With this technique, conductivity distribution variation inside the intracerebral region associated with pathological change can be visualized. However, image reconstruction is a severely ill-posed inverse problem and it possesses inherently soft-field properties. Particularly, irregular and multi-layered head structures along with low-conductivity skulls further aggravate the challenge for accurate reconstruction. To solve this problem, an attention-based deep convolutional neural network (ADCN-Net) is proposed for imaging the conductivity distribution in brain EIT. It employs two branches convolution block (TBC-block), an adaptive feature recalibration (AFR) module, a dense block, and an attention mechanism. Based on this method, the problem of gradient disappearance is mitigated, multi-scale advanced features of voltage sequences can be extracted and the critical features can be captured, thereby improving the quality of reconstruction. To evaluate the performance of the proposed method, extensive simulation work is conducted on a three-layer head model, which includes a simulation of intracerebral hemorrhage (ICH). Besides, phantom experiments are carried out. Both qualitative evaluation and quantitative analysis demonstrate that the proposed method performs well in the reconstruction of ICH under various conditions. In addition, the ablation study further validates the advantage of the proposed method. This work offers an effective alternative for accurate reconstruction in imaging ICH in EIT.