Systemic lupus erythematosus (SLE) is a systemic autoimmune disease affecting multiple organs and systems. Approximately one-third of SLE patients exhibit ocular manifestations, with retinopathy being the most severe due to its potential to cause significant vision loss and even blindness. Thus, early screening, evaluation, and management of SLE-related retinopathy are crucial. Fundus examination is the most commonly used method for observing the retina, allowing for direct visualization of retinal blood vessels and providing early, non-invasive, and quantitative indicators. However, research on retinal changes in pediatric SLE patients remains limited. The subtle nature of retinal vascular alterations in these patients makes detection by the naked eye particularly challenging. Additionally, SLE-related retinopathy is classified as a rare disease, further complicating patient data collection and research efforts. To address these challenges, this study introduces the Bayesian Random Semantic Data Augmentation (BSDA) module, which enhances the semantic features of raw retinal fundus images by combining semantic direction and magnitude. This approach effectively mitigates the issues posed by limited data availability and the subtle variations in retinal vascular characteristics. Experimental results demonstrate that incorporating the BSDA module into the backbone of the classification network significantly improves the model's classification performance in detecting the risk of SLE-related retinopathy, as evidenced by metrics such as ACC, AUC, and F1 scores. This enhancement holds significant clinical value, enabling physicians to more accurately and earlier identify potential retinal changes in pediatric SLE patients, thereby providing a foundation for timely intervention and treatment, ultimately improving patient outcomes.