Electrical resistance tomography （ERT）, as a non-invasive tomographic technique, has broad application prospects in fields such as geological exploration, industrial inspection, and biomedicine. Due to the ill-posedness and nonlinearity inherent in its inverse problem, conventional image reconstruction algorithms struggle to meet the demands of complex operating conditions in terms of imaging accuracy and robustness. In recent years, deep learning-based image reconstruction methods for electrical resistance tomography have demonstrated strong nonlinear mapping capabilities and effectively improved imaging accuracy. However, they still face bottlenecks such as heavy reliance on large-scale labeled training datasets, weak generalization capability, and insufficient noise immunity. To address these issues, a semi-supervised learning MultiU-Net network is proposed for ERT image reconstruction. Built upon a consistency-constrained semi-supervised learning strategy with an encoder-decoder framework, this network reduces dependence on labeled training data while improving generalization and anti-noise performance. A channel attention optimization mechanism is introduced into the feature extraction module to adaptively adjust the weight distribution of feature channels, thereby enhancing training efficiency and feature representation capability, and consequently improving image reconstruction accuracy. Experiments on unlabeled data ratio optimization show that optimal imaging performance is achieved when labeled data constitutes 50% of the training dataset. Noise immunity tests demonstrate that the reconstructed images remains stable under 30 dB Gaussian white noise and exhibits almost no distortion at 50 dB. Results from simulation, static, and dynamic image reconstruction experiments indicate that, compared with the linear back projection (LBP) algorithm, Tikhonov algorithm, Landweber algorithm, and Convolutional Neural Network-Visual Geometry Group algorithm, the MultiU-Net network achieves superior performance in both visual quality and reconstruction accuracy.