Yin Yang1, Yiping Ge1, Lifang Guo1, Qiuju Wu1, Lin Peng2, Erjia Zhang1, Junxiang Xie3, Yong Li3, Tong Lin1. 1. Department of Cosmetic Laser Surgery, Hospital of Skin Diseases, Institute of Dermatology, Chinese Academy of Medical Sciences and Peking Union Medical College, Nanjing, China. 2. Department of dermatology, Tongji University Affiliated Tongji Hospital, Shanghai, China. 3. Institute of Medical Information, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
Abstract
OBJECTIVE: This study used deep learning for diagnosing common, benign hyperpigmentation. METHOD: In this study, two convolutional neural networks were used to identify six pigmentary diseases, and a disease diagnosis model was established. Because the distribution of lesions in the original training picture is very complex, we cropped the image around the lesions, trained the network on the extracted lesion images, and fused the verification results of the overall picture and the extracted picture to assess the model performance in identifying hyperpigmented dermatitis pictures. Finally, we evaluated the image recognition performance of the two convolutional neural networks and the converged networks in the test set through a comparison of the converged network and the physicians' assessments. RESULTS: The AUC of DenseNet-96 for the overall picture was 0.98, whereas the AUC of ResNet-152 was 0.96; therefore, we concluded that DenseNet-96 performed better than ResNet-152. From the AUC, the converged network has the best performance. The converged network model achieved a comprehensive classification performance comparable to that of the doctors. CONCLUSIONS: The diagnostic model for benign, pigmented skin lesions based on convolutional neural networks had a slightly higher overall performance than the skin specialists.
OBJECTIVE: This study used deep learning for diagnosing common, benign hyperpigmentation. METHOD: In this study, two convolutional neural networks were used to identify six pigmentary diseases, and a disease diagnosis model was established. Because the distribution of lesions in the original training picture is very complex, we cropped the image around the lesions, trained the network on the extracted lesion images, and fused the verification results of the overall picture and the extracted picture to assess the model performance in identifying hyperpigmented dermatitis pictures. Finally, we evaluated the image recognition performance of the two convolutional neural networks and the converged networks in the test set through a comparison of the converged network and the physicians' assessments. RESULTS: The AUC of DenseNet-96 for the overall picture was 0.98, whereas the AUC of ResNet-152 was 0.96; therefore, we concluded that DenseNet-96 performed better than ResNet-152. From the AUC, the converged network has the best performance. The converged network model achieved a comprehensive classification performance comparable to that of the doctors. CONCLUSIONS: The diagnostic model for benign, pigmented skin lesions based on convolutional neural networks had a slightly higher overall performance than the skin specialists.