Assessing the severity of Atopic Dermatitis (AD) traditionally relies on face-to-face assessments by healthcare professionals. Such approaches are resource-intensive for participants and staff, challenging during pandemics, and prone to inter- and intra-observer variation. We aim to investigate to what extent computer vision algorithms can help standardise and automate the detection and assessment of AD severity using real-world digital images, without human intervention. We developed EczemaNet, a deep learning computer vision pipeline to detect and assess AD severity from digital camera images. We first trained a model that can detect AD lesions in images using the data provided by four dermatologists who delineated (“segmented”) AD regions in 1345 images from 287 children. We then trained a second model that can assess seven AD disease signs from the AD regions identified. EczemaNet demonstrated good performance for assessing the AD severity in real-world images, while being robust to poor imaging conditions. We noted poor inter-rater reliability in the segmentation of AD regions by dermatologists, i.e. dermatologists rarely reached a consensus on the location of AD lesions in the images. We demonstrated the potential of deep learning for assessing AD severity from digital camera images. Nevertheless, we highlighted the challenge of accurately detecting AD lesions. It may limit the performance of algorithms attempting to assess AD severity from real-world camera images.

Assessing the severity of Atopic Dermatitis (AD) traditionally relies on face-to-face assessments by healthcare professionals. Such approaches are resource-intensive for participants and staff, challenging during pandemics, and prone to inter- and intra-observer variation. 
We aim to investigate to what extent computer vision algorithms can help standardise and automate the detection and assessment of AD severity using real-world digital images, without human intervention. 
We developed EczemaNet, a deep learning computer vision pipeline to detect and assess AD severity from digital camera images. We first trained a model that can detect AD lesions in images using the data provided by four dermatologists who delineated (“segmented”) AD regions in 1345 images from 287 children. We then trained a second model that can assess seven AD disease signs from the AD regions identified. 
EczemaNet demonstrated good performance for assessing the AD severity in real-world images, while being robust to poor imaging conditions. We noted poor inter-rater reliability in the segmentation of AD regions by dermatologists, i.e. dermatologists rarely reached a consensus on the location of AD lesions in the images. 
We demonstrated the potential of deep learning for assessing AD severity from digital camera images. Nevertheless, we highlighted the challenge of accurately detecting AD lesions. It may limit the performance of algorithms attempting to assess AD severity from real-world camera images.

ESDR129 - Fully automated assessment of Atopic Dermatitis severity from real-world digital images

The scientific meeting of the European Society for Dermatological Research (ESDR) is held yearly, usually in September. With an international attendance of more than 1000 skin scientists from around the globe, the four-day event includes short oral presentations from peer-reviewed abstracts, keynote lectures from international experts, sponsored symposia, satellite programs, educational events as well as an important focus on posters, networking and building collaborations amongst scientists. For more information about this year’s meeting visit our website.