Ai-Enabled Acne Diagnosis And Smart Dermatologist Referral System

Abstract

Facial acne is a very common dermatosis with an economic and psychological cost for the person, which makes the proper severity graduation a must for adequate therapeutic performance. In this study, we proposed a new grading schema for acne that accommodates different acne morphologies and introduced a quantitative metric for a proper evaluation of the severity. Faced with the challenge of separating between acne lesions with similar visual properties and their count being critical, we extracted patches of skin from the images of the face and tested a number of deep learning architectures, such as YOLOv8, YOLOv7, YOLOv6, and a conventional convolutional neural network (CNN). YOLOv8 improves the fidelity of images by median filter and histogram equalization, feature representation by channel attention module, and reduces class imbalance by focal loss region. In addition, we conducted model pruning and knowledge distillation using features to decrease the computational overhead. Following inference, the dermatoses were bound to generate boxes around, and the grading improved using ancillary patient metadata. The complete pipeline was implemented on a mobile platform, which results in an application that gives users a dermatologist-level evaluation of acne severity. Comparative evaluation between the four models used, evaluated on the basis of confusion matrix, F1-score curve, Precision-Recall Curve, Recall curve, and Precision Curve, showed better results of YOLOv8 over its counterparts, and it showed better diagnostic performance.