<p>Surface lesion diagnosis in medical and agricultural domains plays a critical role in early disease detection, treatment planning, and yield or health management. In recent years, deep learning has achieved remarkable success in surface lesion analysis, including skin lesion assessment in medical imaging and plant leaf disease recognition in precision agriculture. Nevertheless, diagnostic performance remains constrained by challenges such as complex backgrounds, subtle lesion boundaries, and significant scale variations. To address these issues, attention mechanisms have been increasingly incorporated into deep learning models to selectively emphasize lesion-relevant regions while suppressing irrelevant information. This survey presents a comprehensive review of attention-based deep learning approaches for surface lesion diagnosis across both medical and agricultural applications. We introduce a novel perceptual hierarchy taxonomy, which systematically categorizes existing attention mechanisms into data-aware, model-aware, and task-aware paradigms. Based on this taxonomy, we systematically review state-of-the-art methods, revealing shared methodological principles and domain-specific characteristics. Furthermore, we contextualize these paradigms within the evolution of vision foundation models, summarize widely used datasets and evaluation metrics, and identify promising future research directions. This survey offers a structured and unified reference to facilitate further advances in intelligent surface lesion diagnosis.</p>