198 Reliable and interpretable segmentation for remote assessment of atopic dermatitis severity using digital images

Pan, K., Attar, R., Hurault, G., Williams, H. and Tanaka, R. (2023) 198 Reliable and interpretable segmentation for remote assessment of atopic dermatitis severity using digital images. Journal of Investigative Dermatology, 143 (5), S34. (doi:10.1016/j.jid.2023.03.200).

Abstract

Assessing the severity of atopic dermatitis (AD) traditionally relies on a face-to-face assessment by healthcare professionals and may suffer from inter- and intra-rater variability. With the increasing demand for better patient self-care and the growing popularity of telemedicine post-pandemic, the ability to assess AD severity remotely from digital images, such as those taken with a smartphone camera, is becoming increasingly important. Previously, we developed EczemaNet, a fully automated computer vision pipeline for detecting and assessing AD severity. It demonstrated good performance for assessing AD severity in real-world images while being robust to suboptimal imaging conditions. However, we recognize that it had limitations in practical use due to its lack of interpretability in AD area segmentation and to the need for more reliable AD segmentation data provided by specialists. At the same time, we also found a poor agreement for AD segmentation in digital images, with the average interclass correlation coefficient among four dermatologists being 0.45 on 80 digital images. To address these challenges, we improved EczemaNet pipeline to perform AD segmentation in a more reliable and interpretable fashion. The new pipeline uses pixel-level segmentation and data augmentation to improve the quality and robustness of AD lesion detection. We achieved pixel-level AD segmentation using U-Net architecture and evaluated the reliability of the pipeline using various data augmentation methods such as Pix2Pix. Our investigation found that the use of whole-skin images for model training is a viable alternative as a data collection strategy, which would allow the data acquisition to be more cost-effective without affecting the system's final performance.

This record has no associated files available for download.

Contributors

Download statistics