Automated acne lesion counting from subpar images via memory classifiers

Quantifying acne lesions from photographs is an essential component of acne evaluation. Algorithms are available to automate this process, but they can be easily disrupted by common image noise such as blur or illumination issues. We propose a new automated acne lesions classification method that is capable of withstanding common image corruptions by adapting the memory classifier technique and introducing new expert-defined acne visual features to characterize acne lesions. We then integrate the memory classification algorithm into an end-to-end framework to automatically detect lesions, classify them, and assign an acne severity grade from an image of acne-affected skin. On a dataset consisting of 4658 primary lesions (comedones, papules/pustules, and nodules) taken from 140 lateral facial images from a retrospective de-identified cohort of 63 pediatric and 35 adult patients, the proposed method achieved 87.58% overall classification accuracy. Its performance was similar after adding common noise perturbations to the images. Embedding the classification algorithm within an end-to-end framework to count lesions and grade acne severity from an input image resulted in a mean square error of 0.99 for overall acne severity (against the clinicians’ baseline on an 8-pt severity scale). The results show that the proposed algorithm is both accurate in classifying acne lesion types and capable of withstanding noise perturbations in images. Furthermore, they demonstrate its potential use for automating lesion counting and acne severity grading in research and clinical settings.