Comparative Analysis of AI Models for Atypical Pigmented Facial Lesion Diagnosis

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Diagnosing atypical pigmented facial lesions (aPFLs) is challenging for dermatologists, with visual assessment playing a critical role.
• Inaccurate diagnoses can lead to mismanagement and potential patient harm.
• AI offers potential to enhance diagnostic accuracy and support clinical decision-making in dermatology.

Purpose of the Study:

• To evaluate and compare the effectiveness of machine learning (logistic regression) and deep learning (CNN) models for dermoscopic diagnosis of aPFLs.
• To assess AI's role in managing aPFLs by analyzing diagnostic accuracy against dermatologists' assessments.

Main Methods:

• Analysis of 1197 dermoscopic images of excised facial lesions, classified into seven categories (LM, LMM, AN, PAK, SL, SK, SLK).
• Development and comparison of logistic regression models (lesion features and metadata) and CNN models (image analysis).
• Stratified analysis of dermatologists' diagnostic accuracy compared to AI model performance on a test set.

Main Results:

• Dermatologists achieved 71.2% accuracy differentiating malignant from benign lesions, but only 42.9% for seven specific diagnoses.
• Logistic regression model: 100% sensitivity, 33.9% specificity, 53.6% accuracy.
• CNN model: 58.2% sensitivity, 90.8% specificity, 59.5% accuracy for melanoma diagnosis, outperforming logistic regression in specificity and overall accuracy.

Conclusions:

• AI models, particularly CNNs, demonstrate potential to enhance diagnostic accuracy for complex dermatological conditions like aPFLs.
• Integrating AI with clinical data and evaluating diverse AI approaches is crucial for developing precise and scalable dermatological applications.
• AI shows a critical role in improving patient management and outcomes in dermatology.