Super-Resolution of Low-Resolution CT Images Using Convolutional Autoencoder Networks with Residual Skip Connections

Image Super-Resolution (SR) has emerged as a critical task in various domains, allowing low-resolution (LR) photographs to be improved into their high-resolution (HR) equivalents. This study proposes a novel autoencoder-based deep learning framework for image super-resolution, addressing the trade-off between resolution enhancement and computational efficiency. The proposed model leverages convolutional autoencoders (CAEs) with residual connections to effectively reconstruct HR images while preserving essential structural details. Its lightweight architecture ensures computational feasibility, enabling real-time applications in domains like digital photography, remote sensing, and medical imaging. The framework is designed to upscale LR images of dimensions $128 \times 128$ to HR outputs with up to $4 \times$ magnification, reaching resolutions of $512 \times 512$. This capability allows for flexible applications across various resolution requirements. The model is evaluated on private dataset, including computed tomography (CT) images. Benchmarked against traditional SR techniques such as bicubic and Lanczos interpolation, which are commonly used in CT scan machines, when compared to conventional techniques, the suggested model performs better in terms of the Structural Similarity Index (SSIM) by $\mathbf{1 1. 6 \%}$ and the Peak Signal-to-Noise Ratio (PSNR) by 13.54%. This work highlights the potential of autoencoder-based models in advancing SR technology for further optimization tailored to specific use cases and integration with multi-modal data.