Reliable and Interpretable Segmentation for Remote Assessment of Atopic Dermatitis Severity Using Digital Images

The document discusses the improvement of EczemaNet, a computer vision pipeline for detecting and assessing the severity of atopic dermatitis (AD) from digital images. The original EczemaNet had limitations due to its lack of interpretability in AD area segmentation and the need for more reliable AD segmentation data. To address these issues, the pipeline was enhanced to perform AD segmentation in a more reliable and interpretable manner using pixel-level segmentation and data augmentation. The reliability of the pipeline was evaluated using various data augmentation methods such as Pix2Pix. The study found that using whole-skin images for model training is a cost-effective data collection strategy without affecting the system's final performance. The average interclass correlation coefficient among four dermatologists was 0.45 on 80 digital images, indicating a poor agreement for AD segmentation in digital images.