Automated Detection and Counting of Acne Lesions

In the acne severity evaluation, the quantification of inflammatory lesions on the patients’ face is a fundamental step. The aim of this work is to develop a method that automatically performs acne lesions counting. We acquired high-resolution multi-polarization images of facial skin using the visible spectrum. 1089 patients with acne grading ranging from clear to severe were imaged from one to four times at two weeks interval. Images from five different point of views were acquired from each subject and converted to a standardized planar representation. Experts annotated these planar images creating the ground truth. Patients were subdivided into training (648), validation (111), and holdout test subsets (330). A U-net architecture was trained for 61 epochs to output a gray-scale image where hyperintensities corresponded to predicted location of potential acne lesion. Subsequently a blob detection algorithm was used to perform the counting of predicted lesion on each patient. Results showed that the correlation coefficient between ground truth and automatized counting was 0.65 (confidence interval (C.I.) 0.61 - 0.69). This result suggests a very good agreement between ground truth and automatized counting, indicating that it is indeed feasible to meet this clinical need using deep learning methods.