091 Automated classification of hidradenitis suppurativa severity by convolutional neural network analyses using clinical images

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