With the extensive application of three-dimensional images, including from medical imaging to virtual reality, higher demands on their resolution and detail expression begin to grow. This paper performs some optimization studies for the super-resolution deep learning algorithm on three-dimensional images by taking full advantage of modern network structure optimizations, the design of loss functions, and various data enhancement and preprocessing techniques. Introducing residual networks and attention mechanisms improved the expressiveness and training efficiency of the model, combining mean square error loss, perceptual loss, and adversarial loss optimized the pixel-level accuracy and visual quality of the image, and using various data enhancement methods enhanced the generalization ability and robustness of the model. The results of the optimized model outperform those of the baseline model by a large margin in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) and also present higher image quality with more vivid detail preservation in their visual effects. The optimization strategies developed in this study effectively improve the performance of 3D image super-resolution reconstruction and provide high-quality image support in the related application fields.