Optimizing Acne Severity Detection: A Deep Learning Approach with Electronic Medical Record System Integration and Diverse Image Data

Abstract

Acne vulgaris is the 8th most common skin ailment in the world. Although there are many studies in European countries to predict the severity level of acne, there is a significant gap in predicting South Asian skin texture, which is different due to inherent biological differences such as the thickness of the dermis, complexion, and frequency of skin sensitivity. Therefore, the study aims to address this gap with a deep learning (DL) algorithm based on images from different nationalities with different skin colours and features who have low-resolution images and often contain more than one acne lesions. The modal was deployed as a Progressive Web Application (PWA) and embedded in an Electronic Medical Record (EMR). 1,148 training images and 100 testing images were acquired from several resources and labelled into five main categories: from 1 (Clear) to 5 (Severe). A transfer learning approach was implemented by extracting image features using a ResNet-152 pre-trained model, then a fully connected layer was added and trained to learn the target severity level from labelled images. OpenCV (Facial landmark and One-Eye modal) is used to find facial landmarks and extract key skin patches from the images. To address the spatial sensitivity of CNN models, an existing image rolling augmentation approach was used to help the trained CNN model to generalise better on testing data. Theoretically, it causes acne lesions to appear in more locations in the training images and improves the generalisation of the CNN model on test images. Finally, the model’s performance was evaluated on 100 test images using RMSE concerning a consensus among experts. In this research, we obtained a lower RMSE value (0.37) compared to the previous studies, and severity levels are categorised into five alphabetical values.