Mevorach Lior, Farcomeni Alessio, Pellacani Giovanni, Cantisani Carmen
Dermatology Unit, Department of Clinical Internal Anesthesiological and Cardiovascular Sciences, "Sapienza" University of Rome, 00161 Rome, Italy.
Faculty of Economics, Tor Vergata University of Rome, 00133 Roma, Italy.
Diagnostics (Basel). 2025 Apr 28;15(9):1115. doi: 10.3390/diagnostics15091115.
: This study aims to evaluate and compare the diagnostic accuracy of skin lesion classification among three different classifiers: AI-based image classification, an expert dermatologist, and a non-expert. Given the rising interest in artificial intelligence (AI) within dermatology, it is crucial to assess its performance against human expertise to determine its viability as a reliable diagnostic tool. : This reader study utilized a set of pre-labeled skin lesion images, which were assessed by an AI-based image classification system, an expert dermatologist, and a non-expert. The accuracy of each classifier was measured and compared against the ground truth labels. Statistical analysis was conducted to compare the diagnostic accuracy of the three classifiers. : The AI-based image classification system exhibited high sensitivity (93.59%) and specificity (70.42%) in identifying malignant lesions. The AI model demonstrated similar sensitivity and notably higher specificity compared to the expert dermatologist and non-expert. However, both the expert and non-expert provided valuable diagnostic insights, especially in classifying specific cases like melanoma. The results indicate that AI has the potential to assist dermatologists by providing a second opinion and enhancing diagnostic accuracy. : This study concludes that AI-based image classification systems may serve as a valuable tool in dermatological diagnostics, potentially augmenting the capabilities of dermatologists. However, it is not yet a replacement for expert clinical judgment. Continued improvements and validation in diverse clinical settings are necessary before widespread implementation.
基于人工智能的图像分类、皮肤科专家和非专家。鉴于皮肤科领域对人工智能（AI）的兴趣日益浓厚，评估其与人类专业知识相比的性能，以确定其作为可靠诊断工具的可行性至关重要。
这项读者研究使用了一组预先标记的皮肤病变图像，由基于人工智能的图像分类系统、皮肤科专家和非专家进行评估。测量每个分类器的准确性，并与真实标签进行比较。进行统计分析以比较三种分类器的诊断准确性。
基于人工智能的图像分类系统在识别恶性病变方面表现出高灵敏度（93.59%）和特异性（70.42%）。与皮肤科专家和非专家相比，人工智能模型表现出相似的灵敏度和明显更高的特异性。然而，专家和非专家都提供了有价值的诊断见解，特别是在对黑色素瘤等特定病例进行分类时。结果表明，人工智能有潜力通过提供第二种观点和提高诊断准确性来协助皮肤科医生。
本研究得出结论，基于人工智能的图像分类系统可能成为皮肤科诊断中的一种有价值的工具，有可能增强皮肤科医生的能力。然而，它尚未取代专家的临床判断。在广泛应用之前，需要在不同临床环境中持续改进和验证。
