DermX: An end-to-end framework for explainable automated dermatological diagnosis

Area of Science:

• Artificial Intelligence in Medicine
• Dermatology
• Medical Image Analysis

Background:

• Automating dermatological diagnosis is crucial due to the high prevalence of skin diseases and a shortage of dermatologists.
• Convolutional neural networks (ConvNets) show promise for diagnosis but lack clinical explainability, hindering adoption.
• Current explainability validation methods are subjective and costly.

Purpose of the Study:

• To introduce DermX, an end-to-end framework for explainable automated dermatological diagnosis.
• To develop a clinically-inspired ConvNet that provides interpretable diagnostic explanations.
• To evaluate both diagnostic and explanation performance against dermatologists.

Main Methods:

• Developed DermX, a ConvNet trained on the DermXDB dataset (554 images) with dermatologist-provided diagnoses and explanations.
• Introduced DermX+, an extension of DermX incorporating guided attention training for enhanced explanation maps.
• Assessed diagnostic performance using F1 scores and explanation performance via identification and localization F1 scores.

Main Results:

• DermX and DermX+ achieved near-expert diagnostic performance with F1 scores of 0.79, comparable to dermatologist F1 score of 0.87.
• Explanation identification F1 scores were 0.77 for DermX and 0.79 for DermX+.
• Localization F1 scores were 0.39 for DermX and 0.35 for DermX+, indicating effective explanation localization.

Conclusions:

• Explainability in automated dermatological diagnosis does not compromise predictive accuracy.
• The DermX framework provides expert-inspired explanations for diagnoses, facilitating clinical trust and adoption.
• This approach addresses the critical need for explainable AI in clinical dermatology.