In underwater wireless optical communication (UWOC) systems, the alignment between the laser transmitter and receiver is disrupted by light scattering and imaging angle variations, reducing spot imaging quality, positioning accuracy, and link stability. To overcome these limitations, a deep learning-based light source detection network (LSDNet) is designed for the active link alignment task in UWOC systems. Within the backbone, a locally sparse dynamic graph construction (MDGC) method, guided by multi-dimensional hybrid collaborative attention, is proposed to learn deep representations of underwater optical images, reduce node redundancy, filter out false spots, and suppress scattering. To train and evaluate the model, we construct the UWOC LED light source detection benchmark dataset (UWLED), encompassing 22,770 high-quality images across diverse complex underwater scenarios. Experimental results demonstrate that the proposed LSDNet outperforms other advanced methods, achieving an AP_50^val of 99.2% and the mean center location error (CLE) below 5.41 pixels on the test set, while also exhibiting outstanding robustness under low-light and scattering conditions. Moreover, LSDNet reduces the number of parameters by 30.8% and achieves a 9.2% higher AP_50^val$ compared to Vision Graph Neural Network (ViG). The UWOC system based on LSDNet achieves a bit error rate (BER) at the 10^-8 level over a distance of 35 meters.