Smart imaging and deep learning for objective psoriasis lesion scoring: a scarletred AI proof-of-concept study

The assessment of psoriasis severity requires detailed, and robust lesion segmentation and tissue classification. This proof-of-concept study aimed to automate the classification of psoriasis disease severity on skin lesions using sophisticated deep learning algorithms. Despite remarkable progress in computer-aided diagnostics, there remains a scarcity of effective methodologies for analyzing and categorizing psoriasis. In this study, which was part of a randomized, controlled trial in psoriasis patients (NCT04394936), we proposed a robust pipeline to collect a dedicated, diverse and curated dataset. From 26 patients, we used a total of 152 images of representative psoriasis target plaques acquired with Scarletred® Vision, an innovative dermatologic software platform holds a CE class Im certification as a Medical Device and operates seamlessly through an iOS-App on smartphones. When paired with the Scarletred® Skin-Patch, it effectively standardizes the imaging procedure and facilitates the measurement of 2D/3D aspects, along with the evaluation of color and texture alterations in the skin. As for the modeling aspect, we have trained a deep learning algorithm that employs simultaneously collected target lesion scores derived from experts and manually delineated areas. This algorithm seamlessly integrates into the platform’s proprietary image augmentation pipeline. Our study findings reveal that the proposed model achieves high precision and accuracy across various tasks. This includes lesion segmentation, yielding a dice coefficient of 0.85 whereas for tissue classification, our model achieves a test accuracy and an F-score, of 88%, respectively. Furthermore, our model was able to estimate lesion severity for erythema, scaling and induration attaining test accuracies of 60%, 65% and 83% respectively by implementing a four-class prediction framework that covers grades 0, 1, 2, and 3. Notably, we demonstrated that successful AI classification has been attained using a significantly reduced input dataset compared to prior undertakings in this domain, showcasing a noteworthy advancement in efficiency and its use for prospective rapid AI prototyping in clinical and hybrid trial environments, with the objective of enhancing expert decision making.