Detecting Eczema Areas in Digital Images: An Impossible Task?

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Traditional atopic dermatitis (AD) severity assessment by clinicians has inter- and intra-rater variability.
• Telemedicine and machine learning (ML) offer potential for automated AD severity assessment from digital images.
• Current ML algorithms for AD severity rely on lesion segmentation data from healthcare professionals.

Purpose of the Study:

• To evaluate the reliability of atopic dermatitis lesion segmentation in digital images.
• To quantify inter-rater agreement among dermatologists for AD segmentation.
• To assess the impact of segmentation reliability on ML-based eczema severity assessment.

Main Methods:

• Four dermatologists independently segmented AD lesions in 80 digital images from a clinical trial.
• Inter-rater reliability was calculated using the intraclass correlation coefficient (ICC) at pixel and area levels.
• ICC was assessed across different image resolutions.

Main Results:

• The average ICC for AD segmentation was 0.45 (SE=0.04), indicating poor agreement between raters.
• Inter-rater reliability varied significantly across different images.
• AD lesion segmentation in digital images is highly dependent on the individual dermatologist.

Conclusions:

• The poor reliability of AD segmentation among dermatologists poses a significant limitation for training ML algorithms.
• Data used for training ML models for eczema severity assessment must account for segmentation variability.
• Improving standardization or consensus in AD lesion delineation is crucial for reliable automated assessment.