Abstract Background Artificial intelligence (AI) is a transformative diagnostic tool in dermatology. As the prevalence of skin cancer rises and pressure on health services increases, there is an increasing demand for efficient diagnostic tools. Therefore, it is highly relevant to evaluate the diagnostic abilities of AI tools with a focus on not just the sensitivity, but also the specificity, to reduce unnecessary referrals and skin biopsies for benign skin lesions. Objectives To evaluate the effectiveness of AI tools in diagnosing benign skin lesions, with a primary focus on calculating the sensitivity and specificity of AI tools when analysing suspicious skin lesions. Methods A comprehensive search was conducted on multiple databases, adhering to the PRISMA 2020 guidelines. Studies with recorded sensitivity, specificity and diagnosis were included. Nine studies meeting the inclusion criteria were assessed. Data extracted included type of AI model used, sample size and important diagnostic metrics. Risk of bias was also assessed using the Quality Assessment of Diagnostic Accuracy Studies version 2 (QUADAS-2) framework. Results Across the included studies, AI tools demonstrated a specificity of 70–95% with a pooled specificity of 82%, and a sensitivity ranging from 65% to 96% with a pooled sensitivity of 89% when analysing suspicious skin lesions as malignant vs. benign. Conclusions AI tools possess significant potential to streamline dermatological diagnostics, especially in resource-restricted clinical setups. High sensitivity will minimize false negatives, which is crucial for early detection, and high specificity will allow the use of AI tools to autonomously discharge patients with benign skin lesions. However, there is a need for further optimization and training of AI models on more diverse datasets.
Nirula et al. (Fri,) studied this question.