| Objectives:Discuss the accuracy of the deep learning softwares "Youzhi Pifu" and "Ruifu Shibie" for the diagnosis of skin malignant tumors in actual clinical work,and compare their diagnostic accuracy of skin malignant tumor with professional dermatologists in actual clinical work,so as to provide reference for clinical application.Methods:The patients in the study were patients who attended the Dermatology Department of the Second Affiliated Hospital of Kunming Medical University from December 2019 to December 2020 and were not previously diagnosed with skin malignancies.During the first consultation,the deep learning software "Youzhi Pifu"and "Ruifu Shibie" were used to take pictures of the patient’s skin lesions for identification and diagnosis.Included patients were those who had one or more of the diagnoses by the primary doctor or the other two deep learning softwares that suggested that the patients were diagnosed as skin malignant tumors and voluntarily underwent histopathological examination of lesions to confirm the diagnosis after informed consent.Record the diagnosis of the primary doctor and the two deep learning softwares and the patient’s gender,age,place of residence,occupation,sun protection history,lesions’locations,lesions’size,and rupture conditions.The above data were processed by Excel 2013,and the final histopathological diagnosis results were used as the diagnostic gold standard.SPSS 24.0 was used for statistical analysis.Results:1.If only from the perspective of distinguishing cases into skin malignant tumors and skin non-malignant tumors,the dermatologists’ sensitivity to skin malignant tumors is 96.15%,the specificity is 63.83%,the Youden index is 0.60,and the area under the ROC curve is 0.580(95%CI:0.489-0.668),higher than(z=4.07,P<0.05)deep learning softwares " Youzhi Pifu "(sensitivity is 75.64%,specificity is 40.43%,Youden index is 0.16,the area under the ROC curve Is 0.599,95%CI:0.508-0.686)and " Ruifu Shibie "(sensitivity is 73.08%,specificity is 46.81%,Youden index is 0.20,area under ROC curve is 0.599,95%CI:0.508-0.686).There was no significant difference in the recognition ability of skin malignant tumors between " Youzhi Pifu "and " Ruifu Shibie "(z=0.29,P>0.05).2.After histopathological diagnosis,78 cases were finally diagnosed as skin malignant tumors.The specific diseases were mainly BCC(50 cases,64.10%)and SCC(18 cases,23.08%).Male(41 cases,52.56%),female(37 cases,47.44%),the youngest is 30 years old,the oldest is 92 years old,the average age is 67.03 years,the median age is 70 years old,the occupation is mainly farmers(48 cases,61.54%),in most cases,daily sun protection was not carried out(66 cases,84.61%).The shortest course of the case was less than 1 month(29 days),the longest was 50 years,the average course was 5.32 years,and the median course was 2 years.The skin lesions were mainly located in the head,face and neck(59 cases,75.64%)and limbs(12 cases,15.38%);the smallest length of the skin lesions was 4mm,the largest was 300mm,the average length was 28.28mm,and the median length was 20mm.More than half of the skin lesions were ruptured(53 cases,67.95%).3.The overall diagnosis accuracy rate of dermatologists for skin malignant tumor cases is 84.62%,which is higher than the deep learning software " Youzhi Pifu "(51.28%,χ2=19.90,P<0.05)and "Ruifu Shibie "(48.72%,χ2=22.62),P<0.05),but there was no significant difference in the diagnostic accuracy between "Youzhi Pifu "and " Ruifu Shibie "(χ2=0.10,P>0.05).The diagnostic accuracy rate of dermatologists with senior professional titles is 92.86%,which is higher than that of dermatologists with middle and low professional titles(63.64%,χ2=8.24,P<0.05),but There was no significant difference in the diagnostic accuracy between the dermatologists with middle and low professional titles and the diagnostic accuracy of"Youzhi Pifu " and " Ruifu Shibie"(χ2=1.54,P>0.05).4.Logistic regression analysis of possible risk factors misdiagnosed by dermatologists showed that the final accurate diagnosis for the case of broken skin lesions was 0.08 times that of non-broken skin lesions,with a 95%CI of 0.01-0.86 times(P<0.05).The accurate diagnosis of cases by doctors with senior professional titles was 6.57 times as much as that by doctors with middle and low professional titles,with a 95%CI of 1.44-30.08 times(P<0.05).5.Logistic regression analysis of possible risk factors for misdiagnosis of"Youzhi Pifu" showed that the diagnostic accuracy of cases with daily sun protection was 0.24 times higher than that without sun protection,and 95%CI was 0.06～0.99 times(P<0.05).For every 1mm increase in the lesion length,the cases diagnosis accuracy is 0.97 times of the diagnosis inaccuracy,with a 95%CI of 0.95-1.00 times(P<0.05).6.Logistic regression analysis of possible risk factors for misdiagnosis of "Ruifu Shibie " showed that the diagnostic accuracy of cases with daily sun protection was 0.15 times higher than that without sun protection,and 95%CI was 0.03～0.75 times(P<0.05).For every 1mm increase in the lesion length,the cases diagnosis accuracy is 0.97 times of the diagnosis inaccuracy,with a 95%CI of 0.94-1.00 times(P<0.05).Conclusions:1.On the whole,the dermatologists’ ability to recognize cases as skin malignant tumors and skin non-malignant tumors is stronger than that of deep learning softwares"Youzhi Pifu" and " Ruifu Shibie ".2.On the whole,the diagnostic accuracy rate of dermatologists for skin malignant tumors is higher than deep learning softwares " Youzhi Pifu " and " Ruifu Shibie ".There were differences in diagnostic accuracy among dermatologists with different professional titles.The diagnostic accuracy of dermatologists with senior professional titles was higher than that of dermatologists with middle and low professional titles.At the same time,there was no significant difference in diagnostic accuracy between dermatologists with middle and low professional titles and other two deep learning softwares "Youzhi Pifu" and " Ruifu Shibie ".Deep learning software can become an auxiliary tool in the daily clinical work of dermatologists with middle and low professional titles and doctors in areas lacking professional dermatologists.3.The cases of skin malignant tumors included in this study are mainly elderly farmers who work outdoors and do not have sun protection habits.The lesions are mainly located in the exposed parts of the head,face,neck,limbs,etc.The disease types are mostly BCC and SCC,and the proportion of skin damage is relatively high.The overall situation is similar to the results of other domestic studies.4.The possible risk factors that cause the doctor’s misdiagnosis are the rupture of skin lesions and the doctor’s experience;the possible risk factors that cause the misdiagnosis of the deep learning softwares" Youzhi Pifu " and "Ruifu Shibie" are the daily sun protection history and the long diameter of skin lesion. |
PDF Full Text Request