Hidradenitis suppurativa (HS) is characterized by inflammatory nodules, abscesses and fistulas. Severity can be assessed by several scores requiring detailed, time-consuming and error-prone lesion counts. This study aimed to investigate automated HS-severity classification from clinical images. 777 distinct images from 149 patients were taken using smartphones. Images were ambient-light and size-controlled using the Scarletred® Vision platform. A convolutional neural network (CNN) was used to predict Hurley grades. A mixed input neural network was trained for a granular representation of disease activity (scale 0-49). A U-NET algorithm was implemented for localization of diseased skin. To solve class-imbalances, a synthetic dataset (n=7675, 80% training, 20% test) was generated using a data augmentation approach (scaling, normalization, feature engineering). Training CNNs to distinguish no, mild, moderate and severe disease, provided an overall prediction accuracy of 72% (4-scale, multiclass) and 78% (binary; no/mild vs. moderate/severe). The receiver operating curves revealed an AUC of 0.85 (binary) and 0.84-0.89 (multiclass). The class-wise accuracy for each Hurley grade was calculated with 81-88%. No/mild disease could robustly be distinguished from moderate/severe disease (p<.001). The mixed input model accurately aligned with changes in HS disease activity with a NRMSE value of 0.2618. The UNET algorithm localized HS lesions with a pixel accuracy of 88.1% and test loss of 0.42. The CNNs were able to: 1) Distinguish no/mild from moderate/severe disease, 2) classify disease dynamics over time and 3) identify diseased skin areas. This study breaks new grounds for fast, reliable, reproducible and easy-to-use HS severity assessments.

Hidradenitis suppurativa (HS) is characterized by inflammatory nodules, abscesses and fistulas. Severity can be assessed by several scores requiring detailed, time-consuming and error-prone lesion counts. This study aimed to investigate automated HS-severity classification from clinical images. 777 distinct images from 149 patients were taken using smartphones. Images were ambient-light and size-controlled using the Scarletred® Vision platform. A convolutional neural network (CNN) was used to predict IHS4 grades. A mixed input neural network was trained for a granular representation of disease activity (scale 0-49). A U-NET algorithm was implemented for localization of diseased skin. To solve class-imbalances, a synthetic dataset (n=7675, 80% training, 20% test) was generated using a data augmentation approach (scaling, normalization, feature engineering). Training CNNs to distinguish no, mild, moderate and severe disease, provided an overall prediction accuracy of 72% (4-scale, multiclass) and 78% (binary; no/mild vs. moderate/severe). The receiver operating curves revealed an AUC of 0.85 (binary) and 0.84-0.89 (multiclass). The class-wise accuracy for each IHS4 grade was calculated with 81-88%. No/mild disease could robustly be distinguished from moderate/severe disease (p<.001). The mixed input model accurately aligned with changes in HS disease activity with a NRMSE value of 0.2618. The UNET algorithm localized HS lesions with a pixel accuracy of 88.1% and test loss of 0.42. The CNNs were able to: 1) Distinguish no/mild from moderate/severe disease, 2) classify disease dynamics over time and 3) identify diseased skin areas. This study breaks new grounds for fast, reliable, reproducible and easy-to-use HS severity assessments.

ESDR091 - Automated classification of hidradenitis suppurativa severity by convolutional neural network analyses using clinical 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.