Automatic SCOring of Atopic Dermatitis Using Deep Learning (ASCORAD): A Pilot Study

Authors

Alfonso Medela, Taig Mac Carthy, S. Andy Aguilar Robles, Carlos M. Chiesa-Estomba, and Ramon Grimalt.

What the study evaluated

SCORAD combines several observations to measure atopic-dermatitis severity. This retrospective pilot study tested neural networks for two image-based parts of that process: estimating the affected area and rating six visual signs.

Three dermatologist-annotated image datasets were used. The two larger datasets contained only Fitzpatrick skin types I to III. A third dataset contained 112 images of skin types IV to VI and was used to examine performance on more richly pigmented skin.

Reported results

For the visual-sign task, the model had a relative mean absolute error of 13.0%. For lesion segmentation, it reached an area under the curve of 0.93 on the light-skin test data. Performance was materially lower on the small skin-types IV to VI dataset, including an intersection over union of 0.32 compared with 0.64 on the light-skin test set.

Limits

This was an image-based, retrospective pilot study, not a trial of patient outcomes or a replacement for clinician assessment. The source images came from online dermatology atlases, basic demographic data were missing, and the richly pigmented-skin test set was small. The paper presents ASCORAD as a candidate for further study, not as clinical guidance for assessing an individual patient.

Read full text

Available at https://doi.org/10.1016/j.xjidi.2022.100107.