Research On Facial Acne Grading Assessment Method Based On Convolutional Neural Network

IntroductionRecently,along with the rapid development of convolutional neural network,computer vision has been widely used in the medical industry.Dermatology relies on vision and graphs in diagnosis and treatment,thus computer vision is explored deeply in this field.Previous studies suggested that convolutional neural network performs well in the diagnosis of skin diseases.For instances,the Stanford group confirmed in 2017 that convolutional neural network could work as well as human dermatologists in the discrimination of malignant melanomas and benign nevi,keratinocyte carcinomas and benign seborrheic keratosis.Meanwhile,Acne Vulgaris is a commonly seen skin disorder.However,the lack of a universally approved grading scale has led to misbehavior in treatment and inconvenience in clinical trials.So far,deep learning is rarely used in the acne severity grading.ObjectivesTo apply different convolutional neural network algorithm in the grading of facial acne.To determine the most useful deep learning model in assistance to some specific skin disease assessment.MethodsPresent acne severity grading scales were reviewed in this study.The researcher chose the face to be the evaluation site,and adopted the criterion established by the Japanese dermatologist Hayashi et al.Acne Vulgaris patients were enrolled in the dermatology clinic of Peking Union Medical College Hospital.After oral consent,photographs of both sides of the face were taken by a interchangeable lens digital camera.The image classification and object detection models,simulating grading and counting respectively,were adopted in this acne grading problem.A qualified dermatologist trained with the Hayashi criterion graded all the sample photographs,labeled all the papules and pustules in the sample photographs with MATLAB.Then the sample was randomly divided into the training set and the test set.Two neural networks,ResNet and Faster R-CNN,were trained independently with the training set,and then gave the prediction results of the test set.In the end,the results of two neural networks were analyzed and compared.Results1)1503 photographs were collected altogether to serve as the dataset in this study.550,638,185 and 130 photographs were graded into mild,moderate,severe and very severe ones,respectively.2658 pustules and 13644 papules were labeled from the photographs.2)The sensitivity(recall)of the mild,moderate,severe and very severe level by deep ResNet was 0.78,0.72,0.57 and 0.71,respectively.The AUC of ROC curve for the mild,moderate,severe and very severe level was 0.7707、0.7244、0.6008、0.7185,while the total accuracy was 0.722 for ResNet in the facial acne severity grading.3)Faster R-CNN obtained different performance under different thresholds of the score.The total accuracy ranged from 0.7 to 0.76.When the threshold was 0.4,the total accuracy reached the highest level of 0.759,though the sensitivity of very severe only left 0.04.Conclusions1)The results indicate that ResNet and Faster R-CNN can both perform well in the facial acne severity grading.It is suggested that Convolutional Neural Network could perform much better with the enlargement of the dataset and the advance of computer calculative capabilities.2)Since skin diseases usually have small lesions and high density,the algorithm model of Image Classification is more suitable for severity assessment3)Further study of Convolutional Neural Network can be helpful in the establishment of a universal acne grading scale,the objective evaluation of clinical trials and the promotion of standardized treatment.