人工智能（AI）在黑色素细胞性病变诊断中的应用：纳入ABCD(E)衍生特征的方法的系统综述。
完整摘要 · 中文译文
黑色素瘤的发病率不断上升，早期且准确的检测可以提高生存率。然而，其复杂的诊断带来了挑战，人工智能（AI）有望克服这些挑战，而ABCD(E)规则（不对称性、边缘不规则、颜色多样、直径和演变）是实现黑色素瘤检测的主要筛查模式。本系统综述评估了人工智能（AI）在黑色素瘤诊断方面的进展，特别关注纳入ABCD(E)标准或受其启发的系统。在Cochrane和PRISMA指南的指导下，本综述纳入了在PubMed、LILACS、Latindex和Cochrane Central Register of Controlled Trials平台上索引的实验性或观察性研究。由两位作者独立进行筛选和数据提取。共评估了2,735条记录，最终筛选纳入21项研究，分析了超过90,000张图像/病例。AI系统，特别是CNN和DCNN，展现出有前景的诊断性能，通常达到或超过人类专家评估，在早期黑色素瘤鉴别方面达到90.3%的曲线下面积（AUC）。观察到患者选择偏倚，限制了研究结果的适用性。AI在皮肤病变诊断方面展现出显著且不断发展的能力，日益整合ABCD(E)标准。虽然提供了增强的诊断一致性，但挑战依然存在，包括数据局限性、模型泛化性和可解释性。
完整摘要 · 英文原文
Melanoma has a growing incidence, and early and accurate detection can enhance survival rates. However, its complex diagnoses pose challenges that could be overcome by Artificial Intelligence (AI), and the ABCD(E) rule (Asymmetry, Border irregularity, Color variegation, Diameter, and Evolution) is a major screening pattern enabling melanoma detection. This systematic review assessed advancements in Artificial Intelligence (AI) for melanoma diagnosis, particularly focusing on systems incorporating or informed by ABCD(E) criteria. Guided by Cochrane and PRISMA guidelines, this review included experimental or observational studies indexed on PubMed, LILACS, Latindex, and Cochrane Central Register of Controlled Trials platforms. Independent screening and data extraction were performed by two authors. 2,735 records were assessed and 21 studies were included in the final screening, analyzing over 90,000 images/cases. AI systems, notably CNNs and DCNNs, demonstrated promising diagnostic performance, often matching or exceeding human expert evaluation, achieving an Area Under the Curve (AUC) of 90.3% for early melanoma differentiation. Patient selection bias was observed, limiting applicability of the findings. AI shows significant and evolving capability in skin lesion diagnosis, increasingly integrating ABCD(E) criteria. While offering enhanced diagnostic consistency, challenges persist, including data limitations, model generalizability, and interpretability.