Deconvolution in radio interferometry faces challenges due to incomplete sampling of the visibilities in the spatial frequency domain caused by a limited number of antenna baselines, resulting in an ill-posed inverse problem. Reconstructing dirty images into clean ones is crucial for subsequent scientific analysis. To address these challenges, we propose a U-Net based method that extracts high-level information from the dirty image and reconstructs a clean image by effectively reducing artifacts and sidelobes. The U-Net architecture, consisting of an encoder-decoder structure and skip connections, facilitates the flow of information and preserves spatial details. Using simulated data of radio galaxies, we train our model and evaluate its performance on the testing set. Compared with the CLEAN method and the visibility and image conditioned denoising diffusion probabilistic model, our proposed model can effectively reconstruct both extended sources and faint point sources with higher values in the structural similarity index measure and the peak signal-to-noise ratio. Furthermore, we investigate the impact of noise on the model performance, demonstrating its robustness under varying noise levels.