Due to the inherent local receptive field limitations of conventional convolutional networks, the representation of complex features in electrical impedance tomography (EIT) is often restricted, leading to inaccurate reconstruction of inclusion conductivity parameters and shape distortion caused by the neglect of global distribution information. To address these challenges, this paper proposes an end-to-end deep learning framework, HiRef-EIT, which integrates attention mechanisms and vision Transformers. Built upon a U-Net backbone, the proposed method employs hierarchical feature stacking to construct a multi-scale latent space representation. A multi-head spatial reduction attention（MHRA） module is designed to reduce computational complexity while preserving global modeling capability. Additionally, a skip bottleneck module combining channel attention and coordinate attention in parallel is introduced to capture the latent representations of conductivity distribution and inclusion shape from both the channel and spatial perspectives. A decoder unit based on cross-attention is employed to achieve global-local multi-scale feature fusion, thereby enhancing the model’s representation capacity. HiRef-EIT is trained using a large volume of simulated data to obtain optimal model parameters, and is experimentally validated on diverse complex-shaped phantoms and lung simulation models. Quantitative evaluation metrics demonstrate that the proposed method achieves a root mean square error (RMSE) of 6.028 and a structural similarity index (SSIM) of 0.956, with visual results showing strong consistency with the ground truth distributions and boundaries. The experimental results reveal that HiRef-EIT exhibits excellent robustness and generalization. Compared with traditional convolutional imaging models, the proposed method offers a high-quality imaging solution for EIT, thus holding significant theoretical and practical value for applications in medical diagnostics and industrial process monitoring.