<p>RGB-D salient object detection (SOD) leverages depth cues to improve object localization and boundary accuracy. However, in complex scenes, many RGB-D SOD methods cannot dynamically associate depth features with spatial structures, which weakens foreground–background separation and leads to ambiguous boundaries under cluttered backgrounds. To address this issue, we propose LCDNet, a learnable clustering-driven framework that introduces an end-to-end differentiable clustering layer to explicitly decouple foreground and background semantics and to support reliability-aware cross-modal fusion. First, we design a Clustering Refinement Block that projects depth features into an optimizable clustering space and iteratively updates foreground and background centroids to learn their semantic distributions; CRB then produces soft masks to recalibrate RGB and depth features before fusion. Second, we develop a Gated Bidirectional Attention Module that performs bidirectional cross-attention to align RGB–depth semantics and applies a spatially varying gate to adaptively regulate fusion weights, suppressing unreliable modality responses. Third, we introduce a Multiscale Edge-Aware decoder that injects edge cues during upsampling and aggregates multiscale context to strengthen boundary-aware representations and recover fine details. Extensive experiments on six benchmark datasets (NJU2K, NLPR, STERE, SIP, DUT-RGBD, and DES) show that LCDNet consistently achieves state-of-the-art performance and competitive efficiency, outperforming 12 existing RGB-D SOD methods in comprehensive evaluations.</p>