Detecting objects in underwater scenes is essential for constructing Digital Twin (DT) models tailored to marine settings. These models enable applications like monitoring underwater infrastructure, evaluating ecological conditions, and guiding autonomous underwater systems. However, accurate detection remains challenging due to complex optical distortions and limited visibility inherent in underwater scenes. Existing deep learning methods often struggle to capture the fine-grained contextual relationships required for reliable representation within DT frameworks. In this work, we propose a novel end-to-end architecture that integrates hypergraph based latent space refinement to improve underwater object detection for Digital Twin applications. Our approach begins with a convolutional encoder that projects raw underwater imagery into a latent feature space, where an unsupervised graph is constructed. Leveraging hypergraph convolution, the model enables robust message passing between multiple latent variables simultaneously, enhancing the representation of spatial and semantic relationships. The refined graph is then decoded back into the image space to recover object-level information. We validate the effectiveness of our method through extensive experiments on benchmark underwater datasets, demonstrating superior performance over fourteen state-of-the-art models in preserving object boundaries and extracting critical detail for high-fidelity Digital Twin construction.