Early detection of sexually transmitted infections from skin lesions with deep learning: a systematic review and meta-analysis

Abstract

Background: Sexually transmitted infections (STIs) are a substantial public health concern. We aimed to evaluate the accuracy and applicability of deep learning algorithms in the early detection of STIs from skin lesions. Methods: In this systematic review and meta-analysis, we searched PubMed, Institute of Electrical and Electronics Engineers Xplore, Web of Science, Scopus for studies employing deep learning for classifying clinical skin lesion images of STIs published between Jan 1, 2010, and Dec 31, 2023. Studies that did not include clinical images were excluded. The primary outcome was diagnostic performance, assessed by pooled sensitivity and specificity. We conducted a meta-analysis of the studies providing contingency tables using a unified hierarchical model. We additionally assessed the quality of the studies using modified QUADAS-2 and CheckList for Evaluation of image-based AI Reports in Dermatology (CLEAR Derm) criteria. This study was registered with PROSPERO, CRD42024496966. Findings: Among the 1946 studies identified, we included 101 in our review. The majority of the included studies focused on mpox (91 [88%] of 101 studies), followed by scabies (eight [8%] studies), herpes (four [4%] studies), syphilis (one [1%] study), and molluscum (one [1%] study). A meta-analysis of 55 studies showed that deep learning algorithms had a pooled sensitivity of 0·97 (95% CI 0·95–0·98) and a specificity of 0·99 (0·98–0·99) for mpox, and a sensitivity of 0·95 (0·90–0·98) and specificity of 0·97 (0·86–0·99) for scabies. The majority of studies (86 [85%] of 101 studies) utilised public datasets; traditional convolutional neural networks with backbone architectures such as ResNet and VGGNet were used in all studies. However, notable quality issues related to the data, technical descriptions of labelling methods and diagnostic label references, technical assessment for public evaluation of algorithms, benchmarking and bias assessments, application descriptions of use cases, and target conditions and potential impacts were identified in CLEAR Derm. Potential biases in performance evaluation metrics and applicability concerns in the data, deep learning algorithms, and performance evaluation metrics might impede the generalisability of these models to real-world clinical practice and STI screening across diverse populations. Interpretation: Although deep learning shows potential for early detection of STIs, there are challenges to ensuring the generalisability of such algorithms due to limited heterogeneous data. Standardised, diverse skin lesion image datasets are crucial to ensure fair comparisons and reliable performance. Funding: City University of Hong Kong.