Machine Learning-based Complementary Artificial Intelligence Model for Dermoscopic Diagnosis of Pigmented Skin Lesions in Resource-limited Settings.

Background: Rapid advancements in big data and machine learning have expanded their application in healthcare, introducing sophisticated diagnostics to settings with limited medical resources. Notably, free artificial intelligence (AI) services that require no programming skills are now accessible to healthcare professionals, allowing those in underresourced areas to leverage AI technology. This study aimed to evaluate the potential of these accessible services for diagnosing pigmented skin tumors, underscoring the democratization of advanced medical technologies. Methods: In this experimental diagnostic study, we collected 400 dermoscopic images (100 per tumor type) labeled through supervised learning from pathologically confirmed cases. The images were split into training, validation, and testing datasets (8:1:1 ratio) and uploaded to Vertex AI for model training. Supervised learning was performed using the Google Cloud Platform, Vertex AI, based on pathological diagnoses. The model's performance was assessed using confusion matrices and precision-recall curves. Results: The AI model achieved an average recall rate of 86.3%, precision rate of 87.3%, accuracy of 86.3%, and F1 score of 0.87. Misclassification rates were less than 20% for each category. Accuracy was 80% for malignant melanoma and 100% for both basal cell carcinoma and seborrheic keratosis. Testing on separate cases yielded an accuracy of approximately 70%. Conclusions: The metrics obtained in this study suggest that the model can reliably assist in the diagnostic process, even for practitioners without prior AI expertise. The study demonstrated that free AI tools can accurately classify pigmented skin lesions with minimal expertise, potentially providing high-precision diagnostic support in settings lacking dermatologists.