To address the challenges of distortion and insufficient accuracy in image reconstruction of lung electrical impedance tomography (EIT), a multi-scale dense attention network (MsDA-Net) is proposed in this study to improve the reconstruction accuracy of lung ventilation and lesions based on EIT technology. As a direct estimation framework, MsDA-Net integrates dilated convolution, multi-scale dense connection, and attention mechanism to establish the end-to-end image reconstruction architecture with strong feature representation and reuse capabilities, aimed at fully exploiting deep nonlinear features from voltage measurements to improve the accuracy of lung EIT imaging. Both simulation and mapping model experiments are implemented to comprehensively evaluate the performance of MsDA-Net. Simulation results show that the lung contours and lesion structures can be effectively reconstructed by MsDA-Net. Compared with traditional imaging algorithms, the reconstructed images achieve significant improvement in visual quality and quantitative indicators. The average correlation coefficients (CCs), structure similarity index measures (SSIMs), root mean square errors (RMSEs), and peak signal-to-noise ratios (PSNRs) can reach 0.987 1, 0.903 5, 0.060 5, and 31.671 6 dB, respectively. The accuracy of MsDA-Net is similar to that of the frontier model (two-branch hyper-convolution U-Net and attention-based deep convolution neural network), which further confirms the effectiveness and progressiveness of MsDA-Net. Meanwhile, MsDA-Net shows excellent noise robustness, and the images can still maintain basic usability under 20 dB Gaussian white noise interference. Constructing the mapping models within a circular domain based on lung CT images to validate the practicality of MsDA-Net, the results indicate that the shapes and sizes of targets within the field are more accurately reconstructed by MsDA-Net. As the conductivity distribution within the field becomes more complex, the reconstruction accuracy shows a decreasing trend. However, the average CCs, SSIMs, RMSEs, and PSNRs of the reconstructed images can still reach 0.943 1, 0.857 5, 0.109 6, and 19.392 1 dB, respectively.