<p>Underwater salient object detection (USOD) emerges as a pivotal research area in marine perception and exploration, aiming to achieve pixel-level identification and segmentation of salient objects in underwater environments. Despite notable progress in deep learning applications, existing USOD methodologies remain constrained by the inherent limitations of underwater optical degradation. The coexistence of low-contrast characteristics and non-uniform illumination significantly compromises feature discriminability, while ubiquitous additive noise imposes substantial computational burdens further hindering the fulfillment of real-time operational requirements. To address these challenges, we propose a multi-stage differential-aware attention network (MDANet) for real-time USOD implementation. The salient properties of our MDANet are: (1) a boundary-aware feature aggregation module (BAFAM) is presented, which facilitates the prominence of object regions through multi-dimensional integration of channelwise, spatial, and scale-aware features, (2) a salient feature reconstruction unit (SFRU) is designed, which leverages cross-scale interactions for multi-level feature fusion, enabling effective recovery of fine-grained object details, and (3) a lightweight saliency head adapted to our MDANet is constructed, which generates high-quality saliency maps with minimal computational overhead, ensuring deployment feasibility in time-sensitive underwater applications. Comprehensive experiments on the benchmark USOD10K and SUIM data sets demonstrate that our method outperforms state-of-the-art methods in USOD precision while maintaining superior computational efficiency to meet real-time requirements.</p>