| Part Ⅰ:A web survey study of Chinese dermatologists’ attitudes toward artificial intelligenceBackground:Artificial intelligence(AI)has received increasing attention and application in the medical field.As dermatology for diagnosis based on morphology,it is very suitable for the development of artificial intelligence in dermatology.There have already been many AI researches for skin diseases in our country,and there are many AI products gradually being applied in the daily diagnosis and treatment activities.However,the attitude of Chinese dermatologists toward AI is still unclear,and there is no relevant investigation and research.Objective:To understand Chinese dermatologists’ concerns about AI information through a web-based questionnaire,to target the roles of AI,and to analyze and explore new application scenarios in the field of AI from the perspective of dermatologists,as well as which dermatological diseases are more suitable and need to be developed for AI research and applications.Methods:The online questionnaire was designed by experts from the Chinese Skin Image Database(CSID).The questionnaire was published on the UMER Doctor application,and Chinese dermatologists were invited to participate in the questionnaire survey online.Mann-Whitney U test and Kruskal-Wallis H test were used to compare the differences of dermatologists’ attitudes towards AI in different groups(age,gender,hospital level,education degree,professional title,and hospital ownership).Spearman’s rank correlation test was used to calculate the correlation between stratified factors and dermatologists’ attitude towards AI.All analyses were performed using SPSS(version 22.0)software,and two-sided p values less than 0.05 were considered statistically significant.Results:A total of 1,228 Chinese dermatologists from 30 provinces,autonomous regions,municipalities,and other regions(including Hong Kong,Macao,and Taiwan)participated in this survey and filled in a valid questionnaire.The participating dermatologists obtained AI-related information mainly through the Internet,conferences,or forums and 70.51%of the participating dermatologists obtained AI-related information through two or more approaches.Overall,99.51%of the dermatologists involved in the survey paid attention(general attention,passive attention,and active attention)to information related-AI.Stratified analyses showed that dermatologists of different genders,hospital levels,education degree,and professional title have statistically significant differences in the degree of attention(non-attention,general attention,passive attention,and active attention)related to AI information(p≤1.79E-02).95.36%of the participated dermatologists agreed that the role of AI is to assist dermatologists in the daily diagnosis and treatment activities.The stratified analyses of AI roles(unconcerned,useless,assisting dermatologists,and replace dermatologists)showed that there was no statistically significant difference except for the hospital level(p=4.09E-03).The correlation between stratification factors and the degree of attention of AI and the perspective of artificial intelligence roles is extremely weak.Besides,64.17%of dermatologists thought that the application of AI was most needed in secondary hospitals in China,and 91.78%of participated dermatologists thought that AI should be preferentially applied to skin tumors.Conclusions:Most Chinese dermatologists are interested in AI information and acquired information about AI through multiple approaches.Almost all dermatologists focus on information about AI and think the role of AI is in "assisting the daily diagnosis and treatment activities for dermatologists".The application of AI in the future should be mainly concentrated in skin tumors and secondary hospitals.Part Ⅱ:Diagnostic capacity of skin tumor artificial intelligence-assisted decision-making software in real-world clinical settingsBackground:The Youzhi skin artificial intelligence(AI)software is the first yellow human skin tumor AI-assisted decision-making system developed based on skin tumor data from the Chinese Skin Image Database(CSID).The Youzhi skin AI software demonstrated high diagnostic performance through previous testing on specific datasets.Objective:To explore the diagnostic capacity of the Youzhi skin AI software and dermatologists to identify skin tumor images in real-world clinical settings,and the difference between the two in their diagnostic capability in different modes.Methods:A total of 106 patients who had underwent skin tumor resection at the Department of Dermatology of China-Japan Friendship Hospital from July 2017 to June 2019 were selected and were diagnosed as skin tumors by pathological examination.Using the Youzhi skin AI software and 11 dermatologists with different levels of dermoscopy diagnosis,respectively,the diagnosis results were given by identifying the clinical images and dermoscopy images of 106 patients with skin tumors.The main result is to compare the diagnostic accuracy of the Youzhi skin AI software with that of dermatologists,as well as the results of testing in the laboratory using a specific data set.The secondary results include the sensitivity,specificity,positive predictive value,negative predictive value,F-measure,and Matthews correlation coefficient of the Youzhi skin AI software in the real-world.Results:The diagnostic accuracy of the Youzhi skin AI software in real-world clinical settings was lower than that of the laboratory test data(p<0.001).The output result of the Youzhi skin AI software has good stability after several tests.Through the identification of dermoscopic images,the diagnostic accuracy of the Youzhi skin AI software in the diagnosis of benign and malignant diseases and disease types was higher than that in the identification of clinical images(p=0.008,p=0.016,respectively).Compared with dermatologists,the Youzhi skin AI software had higher accuracy in diagnosing disease types by identifying dermoscopic images(p=0.010).By evaluating the diagnostic performance of dermatologists in different modes(scrambled mode,matched mode),the diagnostic accuracy of disease types in matched mode was significantly higher than that in the scrambled mode(p=0.022).The diagnostic accuracy of dermatologists in the diagnosis of benign and malignant diseases by recognizing dermoscopic images was significantly higher than that by recognizing clinical images(p=0.010).Conclusion:The diagnostic accuracy of the Youzhi skin AI software for skin tumors in real-world clinical settings was not as high as that of using special data sets in the laboratory.However,there was no significant difference between the diagnostic capacity of the Youzhi skin AI software and the average diagnostic capacity of dermatologists.Therefore,the software improves performance by further training and may provide auxiliary diagnostic decisions for primary physicians or inexperienced dermatologists in the future.Part Ⅲ A convolutional neural network-based classification method for facial acne vulgarisBackground:Acne vulgaris is a common chronic sebaceous gland disease,which seriously affects the physical and mental health of adolescents.Correctly identifying the different types of skin lesions of acne is necessary for disease diagnosis,grading,and management.Convolutional neural network(CNN)shows good performance in image recognition and image classification and has potential in acne classification.Objective:Based on the CNN algorithm,the VISIA images of acne vulgaris were learned and trained to classify the different types of skin lesions of acne vulgaris.At the same time,it is compared with the acne classification ability of dermatologists.Methods:The VISIA images of acne vulgaris were sourced from the acne vulgaris special disease database of the Chinese Skin Image Database,and the severity of acne vulgaris was graded according to the Pillsbury 4-level classification method.The VISIA images of the three orientations were annotated using Labelme by eight standardized trained dermatologists.The labels were comedones,papules,nodules,pustules,cysts,scars.The review was performed by 2 dermatologists.Efficientnet-B3 was used as the backbone network,which was trained with 5-fold cross-validation.Precision,recall,F1-Score along with ROC curve,and AUC value were used as indicators to evaluate the performance of the classification model.Some acne lesion images were randomly selected from the test set for classification testing and the degree of confirmation of the classification results by junior dermatologists.The Kruskal-Wallis rank-sum test was used to compare the differences between the means of multiple groups of independent data.One-sample t-test or one-sample Wilcoxon signed-rank test were used to compare the classification ability of dermatologists and the classification model.SPSS(version 24.0)was used for all the analyses,and two-sided p values less than 0.05 were considered statistically significant.Results:A total of 1075 patients’ VISIA images were included,and 19669 annotated images were obtained.The classification model was tested by 5-fold validation and showed an average accuracy of 0.8127 on the classification performance of the test set.The recall rates for six categories of skin lesions(comedones,papules,nodules,pustules,cysts,scars)were 0.8510,0.8735,0.1724,0.7169,0.5789,0.8266,respectively.The recall rates of dermatologists for the six categories of lesions selected randomly were 0.8094,0.6480,0.5800,0.8261,0.5080,0.8931,respectively.The difference in the recall rates of papules,nodules,cysts among dermatologists was statistically significant(p<0.05).The average value of the dermatologists’ confirmatory degree scores for the six categories of lesion results were 8.375,6.560,5.271,8.739,6.640,7.914,respectively.The difference in the confirmatory degree scores among the six categories was statistically significant(p<0.001).The classification model had higher recall rates for comedones and papules than dermatologists,and the difference was statistically significant(p<0.01)Conclusions:The classification performance of the current classification model in nodules and cysts is low,which needs to be further improved through supplementing the training samples and so on.Dermatologists are less confident in classifying papules,nodules,cysts than other types of lesions,while they are less confident in the classification results.Artificial intelligence may become a powerful tool to assist dermatologists in diagnosis and treatment of acne. |
PDF Full Text Request