Related Experiment Video
Updated: Aug 23, 2025
09:32
Resolving Water, Proteins, and Lipids from In Vivo Confocal Raman Spectra of Stratum Corneum through a Chemometric Approach
Published on: September 26, 2019
7.2K
Practical Training Approaches for Discordant Atopic Dermatitis Severity Datasets: Merging Methods With Soft-Label and
IEEE Journal of Biomedical and Health Informatics
|October 31, 2022
Summary
Convolutional neural networks (CNN) effectively grade atopic dermatitis (AD) severity. Training CNN models with multiple expert evaluations and specific data merging techniques significantly improves their performance in assessing AD.
Area of Science:
- Dermatology
- Artificial Intelligence
- Medical Imaging
Background:
- Objective assessment of atopic dermatitis (AD) severity is crucial for effective management.
- Current methods for AD assessment can be subjective, necessitating more reliable tools.
Purpose of the Study:
- To evaluate the performance of convolutional neural networks (CNN) in grading atopic dermatitis severity.
- To compare different CNN training strategies for optimizing AD assessment.
Main Methods:
- Five dermatologists evaluated 9,192 AD images using Investigator's Global Assessment (IGA) and six AD signs.
- CNN models were trained using ensemble vs. integration, hard-label vs. soft-label, and train-set pruning approaches.
- Model performance was assessed using macro-averaged AUROC and F1-score.
Main Results:
- The ensemble-soft-label-pruning model achieved AUROC scores of 0.943 (internal) and 0.927 (external).
- The integration-soft-label-whole dataset model achieved F1-scores of 0.750 (internal) and 0.721 (external).
- CNN models trained on multi-evaluator datasets outperformed those trained on individual evaluator datasets, especially when dermatologist assessments were concordant.
Conclusions:
- CNN models demonstrate strong potential for objective AD severity grading.
- Utilizing datasets labeled by multiple evaluators, soft-label merging, and train-set pruning enhances CNN performance for AD assessment.
- These findings suggest improved AI-driven tools for clinical dermatology practice.
