Machine Learning-based Complementary Artificial Intelligence Model for Dermoscopic Diagnosis of Pigmented Skin...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Dermatology

Background:

• Big data and machine learning advancements are enhancing healthcare diagnostics, especially in resource-limited areas.
• Free, no-code artificial intelligence (AI) services are now available for healthcare professionals.
• These tools aim to democratize access to sophisticated medical technologies.

Purpose of the Study:

• To evaluate the potential of accessible AI services for diagnosing pigmented skin tumors.
• To assess the accuracy and reliability of AI in classifying skin lesions.
• To demonstrate the feasibility of using AI for diagnostic support by practitioners without AI expertise.

Main Methods:

• Collected 400 dermoscopic images of pigmented skin tumors (100 per type) from pathologically confirmed cases.
• Utilized supervised learning on the Google Cloud Platform (Vertex AI) with an 8:1:1 training, validation, and testing data split.
• Assessed model performance using confusion matrices and precision-recall curves.

Main Results:

• The AI model achieved an average recall of 86.3%, precision of 87.3%, and accuracy of 86.3%.
• Specific tumor accuracies included 80% for malignant melanoma and 100% for basal cell carcinoma and seborrheic keratosis.
• Testing on separate cases yielded an overall accuracy of approximately 70%.

Conclusions:

• The AI model demonstrates reliable assistance in diagnosing pigmented skin lesions, even for users without AI expertise.
• Free AI tools can accurately classify these lesions, offering high-precision diagnostic support.
• This technology has the potential to significantly aid healthcare settings lacking specialized dermatologists.