Acne detection and severity evaluation with interpretable convolutional neural network models.

Background: Acne vulgaris is one of the most prevalent skin conditions, which harms not only the patients' physiological conditions, but also their mental health. Early diagnosis and accurate continuous self-monitoring could help control and alleviate their discomfort.

Objective: We focus on the development and comparison of deep learning models for locating acne lesions on facial images, thus making estimations on the acne severity on faces via medical criterion.

Methods: Different from most existing literature on facial acne analysis, the considered models in this study are object detection models with convolutional neural network (CNN) as backbone and has better interpretability. Thus, they produce more credible results of acne detection and facial acne severity evaluation.

Results: Experiments with real data validate the effectiveness of these models. The highest mean average precision (mAP) is 0.536 on an open source dataset. Corresponding error of acne lesion counting can be as low as 0.43 ± 6.65 on this dataset.

Conclusions: The presented models have been released to public via deployed as a freely accessible WeChat applet service, which provides continuous out-of-hospital self-monitoring to patients. This also aids the dermatologists to track the progress of this disease and to assess the effectiveness of treatment.