Deep Learning-Based Multiclass Framework for Real-Time Melasma Severity Classification: Clinical Image Analysis and...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Melasma is a common, recurrent skin pigmentation disorder with subjective and variable assessment methods.
• Current clinical assessment tools like the Melasma Area and Severity Index (MASI) suffer from inter-observer variability.
• Objective and consistent assessment of melasma severity is crucial for effective treatment and management.

Purpose of the Study:

• To develop and validate an AI-assisted, real-time framework for classifying melasma severity.
• To leverage deep learning and clinical facial images for objective melasma assessment.
• To improve diagnostic consistency and reliability in melasma evaluation.

Main Methods:

• Collected and preprocessed 1368 anonymized facial images from melasma patients.
• Trained and evaluated six Convolutional Neural Network (CNN) architectures using PyTorch.
• Assessed model performance using accuracy, precision, recall, F1-score, AUC, and Layer-wise Relevance Propagation (LRP) for interpretability.

Main Results:

• GoogLeNet achieved the highest performance with an accuracy of 0.755 and an F1-score of 0.756.
• Area Under the Curve (AUC) values demonstrated strong performance across severity levels: 0.93 (mild), 0.86 (moderate), and 0.94 (severe).
• Layer-wise Relevance Propagation (LRP) analysis confirmed GoogLeNet's robust feature attribution for melasma classification.

Conclusions:

• A robust and interpretable deep learning model for melasma severity classification was developed.
• The AI framework offers enhanced diagnostic consistency, addressing limitations of subjective assessment.
• Future research will explore multimodal data integration for a more comprehensive melasma evaluation.