Deep Learning Algorithms for the Detection of Suspicious Pigmented Skin Lesions in Primary Care Settings: A Systematic Review and Meta-Analysis

• Published in: Curēus (Palo Alto, CA) CA), 2024-07, Vol.16 (7), p.e65122
• Main Authors: Abdalla, Ahmed R, Hageen, Ahmed W, Saleh, Haneen H, Al-Azzawi, Omar, Ghalab, Mahmoud, Harraz, Amani, Eldoqsh, Bola S, Elawady, Fatma E, Alhammadi, Ayman H, Elmorsy, Hesham Hassan, Jano, Majd, Elmasry, Mohamed, Bahbah, Eshak I, Elgebaly, Ahmed
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial intelligence; Bias; Deep learning; Dermatology; Machine learning; Melanoma; Meta-analysis; Neural networks; Oncology; Patients; Primary care; Principal components analysis; Skin cancer; Systematic review; Vision systems
• ISSN: 2168-8184; 2168-8184
• DOI: 10.7759/cureus.65122

Early detection of suspicious pigmented skin lesions is crucial for improving the outcomes and survival rates of skin cancers. However, the accuracy of clinical diagnosis by primary care physicians (PCPs) is suboptimal, leading to unnecessary referrals and biopsies. In recent years, deep learning (DL) algorithms have shown promising results in the automated detection and classification of skin lesions. This systematic review and meta-analysis aimed to evaluate the diagnostic performance of DL algorithms for the detection of suspicious pigmented skin lesions in primary care settings. A comprehensive literature search was conducted using electronic databases, including PubMed, Scopus, IEEE Xplore, Cochrane Central Register of Controlled Trials (CENTRAL), and Web of Science. Data from eligible studies were extracted, including study characteristics, sample size, algorithm type, sensitivity, specificity, diagnostic odds ratio (DOR), positive likelihood ratio (PLR), negative likelihood ratio (NLR), and receiver operating characteristic curve analysis. Three studies were included. The results showed that DL algorithms had a high sensitivity (90%, 95% CI: 90-91%) and specificity (85%, 95% CI: 84-86%) for detecting suspicious pigmented skin lesions in primary care settings. Significant heterogeneity was observed in both sensitivity (p = 0.0062, I = 80.3%) and specificity (p < 0.001, I = 98.8%). The analysis of DOR and PLR further demonstrated the strong diagnostic performance of DL algorithms. The DOR was 26.39, indicating a strong overall diagnostic performance of DL algorithms. The PLR was 4.30, highlighting the ability of these algorithms to influence diagnostic outcomes positively. The NLR was 0.16, indicating that a negative test result decreased the odds of misdiagnosis. The area under the curve of DL algorithms was 0.95, indicating excellent discriminative ability in distinguishing between benign and malignant pigmented skin lesions. DL algorithms have the potential to significantly improve the detection of suspicious pigmented skin lesions in primary care settings. Our analysis showed that DL exhibited promising performance in the early detection of suspicious pigmented skin lesions. However, further studies are needed.