A convolutional neural network trained with dermoscopic images of psoriasis performed on par with 230 dermatologists. (December 2021)

• Record Type: Journal Article
• Authors: Yang, Yiguang; Wang, Juncheng; Xie, Fengying; Liu, Jie; Shu, Chang; Wang, Yukun; Zheng, Yushan; Zhang, Haopeng
• Is Part Of: Computers in biology and medicine. Volume 139(2021)
• Publication Date: 2021-12
• Subjects: Deep-learning -- Convolutional neural networks -- Dermoscopic images -- Papulosquamous skin diseases -- Psoriasis; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2021.104924
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Background: Psoriasis is a common chronic inflammatory skin disease that causes physical and psychological burden to patients. A Convolutional Neural Network (CNN) focused on dermoscopic images would substantially aid the classification and increase the accuracy of diagnosis of psoriasis. Objectives: This study aimed to train an efficient deep-learning network to recognize dermoscopic images of psoriasis (and other papulosquamous diseases), improving the accuracy of the diagnosis of psoriasis. Methods: EfficientNet-B4 architecture was trained with 7033 dermoscopic images from 1166 patients collected from the Department of Dermatology, Peking Union Medical College Hospital (China). We performed a five-fold cross-validation on the training set to compare the classification performance of EfficientNet-B4 over different networks commonly used in previous studies. From the test set, 90 images were used to compare the performance between our four-class model and that of board-certified dermatologists, whose diagnoses and information (e.g., age, titles) were obtained through an online questionnaire. Results: The mean sensitivity and specificity of EfficientNet-B4 on the training set was 0.927 ± 0.028 and 0.827 ± 0.043 for the two-class task, and 0.889 ± 0.014 and 0.968 ± 0.004 four-class task. The diagnostic sensitivity and specificity of the 230 dermatologists were 0.688 and 0.903 for psoriasis, 0.677 and 0.838 for eczema, 0.669 and 0.953 for lichen planus, and 0.832 and 0.932 for the "others" group, respectively; the diagnostic sensitivity and specificity of our four-class CNN was 0.929 and 0.952 for psoriasis, 0.773 and 0.926 for eczema, 0.933 and 0.960 for lichen planus, and 0.840 and 0.985 for the "others" group, respectively. Both the 230 dermatologists and CNN achieved at least moderate consistency with the reference standard, and there was no significant difference between them (P > 0.05). Conclusions: The two-classification and four-classification models of psoriasis established in our study could accurately classify papulosquamous skin diseases. They showed generally comparable performances to the average level of dermatologists and would provide a strong support for the diagnosis of psoriasis. Highlights: Psoriasis is similar to other papulosquamous skin diseases and dermoscopy is a useful tool in the diagnosis of them. Few researches on artificial intelligence have been carried out for recognizing psoriasis based on dermoscopic images. An EfficientNet-based convolutional neural network for classifying psoriasis and other papulosquamous diseases is described. EfficientNet performs better on classifying papulosquamous diseases than other CNNs and on par with 230 dermatologists. EfficientNet has a potential for recognizing atypical dermoscopic features of psoriasis.