Automated Detection, Extraction and Counting of Acne Lesions for Automatic Evaluation and Tracking of Acne Severity

Acne Vulgaris is the most common skin disease and affects 85% of population at some point in life, typically in adolescence. Objective evaluation of acne severity is necessary to asses the efficacy of medical treatment procedures. Traditionally acne evaluation is done by dermatologists manually counting the number of acne lesions through visual inspection or scanning acquired images of the patient's skin. This method is time consuming and requires excessive effort by the physician. In this paper a prototype application for automatic acne detection, lesion counting and reporting through the processing of distance picture taken by mobile devices is developed. The pipeline of the application is composed of body part detection, skin segmentation, heat-mapping, acne extraction and blob detection. Body part detection is accomplished using different Haar-Cascade classifiers; frontal face, right profile, left profile and torso are discriminated. Skin segmentation has been performed using an ensemble of random forest models trained on a augmented version of the FSD dataset. The set of features was engineered combining colour, texture, spatial, shape and unsupervised descriptors and selected by a feature importance step. The a* channel of the CIELab colour space was used for enhance the visual contrast between inflamed zones and healthy skin. Finally acne extraction and blob detection were accomplished through Adaptive Thresholding and Laplacian of Gaussian filtering. Reports are generated containing number, position and ray size of the detected spots.