INTRODUCTION
The measurement and grading of acne severity is a recognized challenge1-3, since it is known to depend on the number and type of lesions and the extent of involvement. The method of reference for acne grading is lesion counting4-6; however, it is time-consuming and not well-suited to clinical practice. For this reason, health agencies4,6 recommend the use of an Investigator Global Assessment (IGA) with up to six ordinal grades (clear, almost clear, mild, moderate, severe and very severe). In order to improve reliability, the FDA recommends co-primary endpoints that evaluate both IGA scale and acne lesion count assessments.6 Computational methods based on imaging techniques7-9 have been proposed in scientific literature to reduce the problems associated to low inter- and intra-rater reliability of acne grading.10-12 To date, however, all attempts have focused on the automatization of lesion counting. In the present work, we intend to investigate if an Artificial Intelligence (AI) could perform IGA grading of digital images of acne patients with reliabilities comparable or superior to those of expert physicians.
METHODS
An imaging device (FaceAtlas, www.linkverse.com) was used to acquire high-resolution multi-polarization images of facial skin using the visible spectrum (Figure 1). During the development of the device, a database of 479 subjects (age range, 11-46 years, median 17.4 years; female 58%) with acne grading ranging from clear to severe, sampled from three ethnic groups (Caucasian-white 79%, African-black 13%, Asian 8%), was acquired at the Department of Dermatology of University “Magna Graecia” after informed consent and approval from the local Institutional Review Board. Three experienced dermatologists assessed acne severity according to the FDA ordinal scale (proportion in the dataset; 0-clear: 18.1%; 1-almost clear: 18.4%; 2-mild: 20.4%; 3-moderate: 23.3%; 4-severe: 19.8%). The clinical evaluation resulting from majority voting among the three raters represented the ground truth for supervised learning.The image acquisition and postprocessing workflow is depicted in Figure 1. As result of this workflow, multiple images collected from a patient were first converted to a standardized planar representation. Subsequently, the planar image was