Evaluation of the area subscore of the Palmoplantar Pustulosis Area and Severity Index using an attention U-net deep learning algorithm.

Palmoplantar pustulosis (PPP) is a variant of pustular psoriasis involving the palms and soles. The severity of PPP is usually evaluated using the Palmoplantar Pustulosis Area and Severity Index (PPPASI). Among the components of the PPPASI, the area of the involved lesion is evaluated differently by raters, who generally make a rough estimate using the eye and not through a specific calculation. To overcome inconsistent evaluation of the area subscore of PPPASI by human raters, we developed and validated deep-learning-based algorithms to enable automated and reliable assessment of the area involved in PPP to provide clinical advantages. In this study, we developed a dataset of 611 images of the palms and soles of 153 patients with PPP. We evaluated the area of the lesion by dividing the number of pixels in the area involved in PPP by the number of pixels in the area of the palms or soles. Using attention U-net, we developed two convolutional neural network (CNN) models that can evaluate the percentage of the affected area (%) and subsequently assign a score ranging from 0 to 6. The area subscore of PPPASI evaluated by the deep-learning algorithm was same or differed by 1-point from the subscore of ground truth in 98.8% of the images. The intraclass correlation coefficient between the CNN and ground truth was 0.879, indicating good agreement. The accuracy and mean absolute error of the model were 66.7% and 0.344, respectively. In a Bland-Altman plot, most of the differences in the percentage of the affected area lay between the 95% confidence interval with a mean difference of 0 and a standard deviation of 0.2. The deep-learning algorithm can provide several clinical advantages by objectively evaluating the components of the PPPASI without concern for disagreement between clinicians. The algorithms further enable cumulative clinical data acquisition related to PPP severity.