| Objective: To construct an artificial intelligence(AI)method based on traditional manual features and explore the value of its automatic identification of fetal facial ultrasound standard plane(FFUSP).Materials and methods: The standard set,validation set and experimental set images of FFUSP at 20～24 weeks of gestation were took as the research object in the study.The standard set images were collected by fetal ultrasound experts of our hospital(grade A hospital)according to relevant guidelines,and were divided into training set and test set with an 8:2 ratio.The validation set images were collected as the same way of the standard set from three tertiary hospitals in the region.The experimental set images were collected from the ultrasonic graphic workstation of our hospital.The three sets all included 3 types of standard plane(nasolabial coronal plane(NCP),median sagittal plane(MSP)and ocular axial plane(OAP))and non-standard plane(N-SP).An AI model based on traditional feature recognition technology was constructed,and the standard set was used to train its ability to recognize and classify NCP,MSP,OAP and N-SP and classification test was performed with test set images.The validation set was used to test the generalization ability of the model for FFUSP images recognition after the AI model built.The experimental set images were classified divided into NCP,MSP,OAP and N-SP by experts.And then AI,junior doctors group(JDG)made up of three doctors with only experience of standardized training for residents and intermediate doctors group(IDG)made up of three doctors with experience of specialized training in obstetric ultrasound classified experiment set images respectively.Observed the AI and the two groups of doctors’ classification accuracy of the experimental set images,and analyzed the difference in the recognition and classification ability of FFUSP between AI and ultrasound doctors with different experience.Results: The standard set contained 1,906 images of FFUSP,including 232,323,456,895 images of OAP,MSP,NCP and N-SP respectively.The images were divided into training set and test sets according to the ratio of 8:2 for each plane.The validation set contained 810 images of FFUSP,including 238,307,122,143 images of OAP,MSP,NCP and N-SP respectively.The experimental set contained 2419 images,including 397,533,594,895 images of OAP,MSP,NCP and N-SP respectively based on expert`s judgment.⒈ AI model had very strong consistency with expert in image recognition and classification of test set(all P<0.05).The sensitivity(SE),specificity(SP),positive predictive value(PPV),negative predictive value(NPV),F1 score(F1),accuracy(ACC),and k value of each plane were respectively:(1)NCP: 93.4%,99.3%,97.7%,98.0%,0.955,97.9%,0.942;(2)MSP: 93.8%,98.8%,93.8%,98.8%,0.938,97.9%,0.925;(3)OAP: 91.3%,98.5%,89.4%,98.8%,0.903,97.6%,0.890.⒉ AI model had strong consistency with expert in image recognition and classification of validation set(all P<0.05).The SE,SP,PPV,NPV,F1,ACC and k value of each plane were respectively:(1)NCP: 86.6%,91.4%,80.8%,94.2%,0.836,90.0%,0.764;(2)MSP: 88.3%,81.5%,74.5%,91.9%,0.808,84.7%,0.673;(3)OAP:88.5%,87.9%,56.5%,97.7%,0.690,88.0%,0.620.⒊ The SE,SP,PPV,NPV and ACC of the experimental set classified by JDG were respectively:(1)NCP:78.6%、86.5%、54.4%、95.2%、85.2%,(2)MSP:79.7%、85.9%、43.5%、96.9%、85.1%,(3)OAP:87.1%、93.1%、62.7%、98.2%、92.4%.The SE,SP,PPV,NPV and ACC of the experimental set classified by IDG were respectively:(1)NCP:88.6%、91.1%、70.9%、97.0%、90.6%,(2)MSP:89.6%、91.2%、66.4%、97.8%、90.9%,(3)OAP:92.7%、94.5%、70.5%、98.9%、94.3%.Intermediate doctors were superior to junior doctors in the recognition and classification of all FFUSP(all P<0.05).The SE,SP,PPV,NPV and ACC of the experimental set classified by AI were respectively:(1)NCP:98.7%、97.1%、90.9%、99.6%、97.5%,(2)MSP:98.1%、96.8%、88.4%、99.5%、97.1%,(3)OAP:94.8%、95.7% 、 77.6% 、 99.2% 、 95.6%.AI`s recognition and classification level of experimental set had a strong consistency with expert`s classification(P<0.05).The sensitivity and specificity of AI to the classification of the three standard planes were all better than those of junior doctors(all P<0.05).The sensitivity and specificity of AI to the classification of NCP and MSP were better than those of intermediate doctors(all P<0.05),while AI was more sensitive to OAP classification than intermediate doctors,and there was no statistically significant difference in SP between the two(P=0.063).AUC of FFUSP classification showed that AI>Intermediate doctor>Junior doctors(all P<0.05).The time of AI classification experimental set was 0.21±0.02 seconds,and the time of expert classification was2.73±0.09 seconds.The efficiency of AI image classification of FFUSP experimental set was significantly better than that of doctors’ manual recognition(P<0.001)Conclusion:1.Specialized training in obstetric ultrasound is helpful to improve doctors’ ability to recognize and classify FFUSP.2.The AI model based on traditional manual features has high efficiency and accuracy in FFUSP recognition and classification,and it has strong ability to recognize the images from different mechanism,so it can be used as an alternative tool for extensive quality control of fetal ultrasound.3.The AI model adopted in this study is more accurate in FFUSP recognition than middle and low level doctors,so it can be used as an auxiliary tool for fetal ultrasound standardized training。... |
PDF Full Text Request