| Thyroid cancer is the most common malignant tumor in endocrine system,accounting for about 90%.In recent years,thyroid cancer has been increasing all over the world,of which papillary thyroid carcinoma(PTC)accounts for more than85%,which is the most common clinical type,endangers people’s health.Because thyroid is a superficial organ,high frequency ultrasound is the first choice for clinical examination,which has the advantages of convenient operation,real-time,noninvasive and good repeatability.However,ultrasound examination has its limitations.The accuracy of manual interpretation is easily affected by the subjective factors of the operator,image artifacts,noise and other objective factors.In addition,the number of patients is large and the number of ultrasound doctors is limited,so the efficiency of manual screening is low.Therefore,the establishment of a thyroid cancer ultrasound image automatic recognition and diagnosis system will have great clinical value.Artificial intelligence(AI)is an important scientific and technological topic in the 21 st century.In recent years,the research on combination of AI and clinical diagnosis and treatment has emerged,especially in the field of medical image recognition.Among them,computer aided diagnosis(CAD)is to process the big data of physiology,biochemistry and medical image of patients by means of machine learning and artificial intelligence analysis and calculation.CAD plays a role in image recognition and depth learning.In this study,we have used several common image feature extraction algorithms,such as local binary mode(LBP),Hessian matrix,direction gradient histogram(HOG),Canny operator and GLCM to extract the features of PTC ultrasonic images,and then set up data sets,and then use SVM,naive Bayesian and K nearest(KNN)and so on A variety of popular classifiers are used to analyze the operation,and then a high precision disease image prediction model is established.Therefore,the purpose of this study is to establish and improve the predictive diagnosis model of thyroid papillary carcinoma based on ultrasound image and artificial intelligence computer-aided diagnosis technology,And try to make a preliminary design and implementation of PTC intelligent identification system,which will lay a foundation for comprehensively improving the clinical efficiency and realizing the AI imaging diagnosis of tumor in the future.In this study,we collected ultrasound images of 101 PTC patients(113 nodules)with pathologically confirmed and selected the same number of normal people as negative control group.The nodules were then evaluated using the American radiological Association thyroid imaging reporting and data system(ACR Ti-rads)grading guidelines.Different feature algorithms are used to extract the image features from the region of interest(ROI)in the longitudinal and transverse section of ultrasound images,and the data set is established.Then a variety of classifiers are selected to evaluate the PTC prediction performance,and the accuracy(ACC)is calculated by comparing with the actual sample situation.Finally,the best prediction model is selected.In addition,269 cases of thyroid papillary carcinoma confirmed by pathology in the last three years were added to verify the model experiment,and good results were achieved.The results are as follows:1.Establish the ultrasonic imaging diagnosis model of thyroid papillary carcinoma by LBP feature algorithm(1)Statistics of basic information of PTC patients in this research: Among 101 patients,there were statistically different categories: gender,24 males(23.76%),77females(76.24%);cervical lymph node metastasis: the metastasis rate of the group ≥45 years old was 16.33%,and the metastasis rate of the 45 year old group was75.00%,of which 29 cases only transferred to the central group lymph nodes,2cases only transferred to the cervical lymph nodes;The metastasis rate of multifocal PTC patients was 69.23%,and that of single focus patients was 43.18%;the metastasis rate of non micro carcinoma(nodule with diameter ≥ 1cm)was 61.76%.The lymph node metastasis rate of the lesions with microcalcification was 67.86%,and 28.07% in the patients without micro calcification.Among these nodules,92cases(81.42%)with A/T≥1 were observed on the transverse plane,21 cases(18.58%)with A/T< 1,In longitudinal plane,59 cases(52.21%)had a A/T ≥ 1 and54 cases(47.79%)had A/T < 1.(2)PTC prediction performance evaluation of four classifiers: the cross section of PTC image is analyzed,and four popular classifiers,support vector machine(SVM),k nearest neighbor(KNN),naive Bayes(NBayes)and decision tree(Dtree),are selected to evaluate the effect.When seven LBP features are used,SVM has the best accuracy,Acc= 9657;when 56 LBP features are used,KNN has the best accuracy,Acc = 0.8436.(3)Evaluation of other classification algorithms: in the relief f algorithm,selectkbest algorithm and support vector machine recursive feature elimination(SVM-RFE)algorithm,the cross section SVM-RFE algorithm has the highest classification prediction accuracy,Acc=0.9655.The results show that all the classifiers perform best in the cross section,and the accuracy of GBC is the highest when 7 features are extracted,with Acc = 0.9655;the accuracy of Lasso is the highest when 28 features are extracted,with Acc = 0.8276.(4)PTC ultrasound image establishment of prediction model on a single section:considering the accuracy of classification prediction and the number of image features,the best prediction model is: on the ultrasonic cross section,SVM classifier is used to simulate the data set,and then 5-fold cross validation method is used to evaluate the effectiveness.When 17 compressed LBP features are used,the classification prediction effect is the best,and the accuracy can reach Acc = 0.9829.2.Integrate five feature extraction algorithms to establish the best prediction model of thyroid papillary carcinoma(1)Five feature extraction algorithms were used to analyze the comparative evaluation of ultrasound images of PTC patients: the prediction performance was evaluated by using three common classifiers: local binary mode(LBP),Hessian matrix(Hessian),direction gradient histogram(HOG),Canny operator and GLCM.In the longitudinal section of ultrasonic,SVM has the best effect on hessian feature(Acc=0.9825);in the cross section image,KNN is the best,and the prediction accuracy of PTC of five feature types is above 0.9000,and the average accuracy is acc=0.9410.(2)The classification prediction is carried out by using deep learning network:the accuracy of the GLCM feature extracted by the deep learning network Le Net-5is the best,Acc=0.9482.(3)Compared with the shallow neural network and four classifiers: the SNN and NBayes,KNN and SVM have the best performance,the longitudinal section Acc=0.9876,and the transverse plane Acc=0.9885.Both of the best models are based on hessian characteristics.(4)The integration analysis of ultrasonic longitudinal and transverse section images: the best prediction model of PTC was obtained by using Hessian and LBP features of two types of ultrasound images and FS+C strategy to select the features.The accuracy of acc=0.9949 was obtained.3.Validation experiment on ultrasonic image prediction model of thyroid papillary carcinomaIn order to further verify the main conclusions in the paper and confirm the clinical practicability of the prediction model,269 cases of thyroid papillary carcinoma with pathologically confirmed and 59 normal control groups collected from 2019 to 2021 were added in this study.The main conclusions are as follows:(1)After the number of feature points is greater than 3,the prediction accuracy of the model using transverse features is mainly distributed between [0.85,0.95],while the accuracy of the model using longitudinal features is mainly distributed between [0.7,0.85].It shows that for the single section prediction model,the conclusion in Chapter 2 is verified that the performance of ultrasonic cross-section feature is better than that of longitudinal section feature.(2)After adding images of different years,sizes and types,the accuracy of the model in this study is still stable,and the error is controlled between [0.018,0.023].After integrating LBP and Hessian features,the model with FS + C strategy is the best,and the ACC reaches 0.971.The main conclusion in Chapter 3 the prediction model of thyroid papillary carcinoma is verified.The two-way ultrasound image differential analysis is carried out by integrating Hessian and LBP features,and the feature is selected by FS + C method,which has the highest prediction accuracy.4.Preliminary design and implementation of PTC intelligent detection systemThrough the research on the methodology of the above PTC best prediction model,we will try to make a preliminary design and application of the system visual interface.The applet is flexible,fast and intuitive,which is convenient for clinicians to use.Verified by random samples,the positive probability of PTC sonogram prediction is high.The probability of negative prediction for normal glands is high,indicating that this procedure has clinical auxiliary value.Future research will continue to expand the scope of diseases,design a variety of core technology algorithms,and optimize the identification system,which can not only ensure its overall safety and stability,but also have flexibility and scalability.The conclusions are as follows:(1)Statistical analysis of PTC patients data sets showed that the incidence rate of female patients was significantly higher than that of men,and those with age less than 45 years,multifocal lesions,microcalcifications in nodules,and non microcarcinoma(diameter greater than 1cm nodules)were all high-risk factors for cervical lymph node metastasis.PTC lymph nodes were easy to metastasize to the central group lymph nodes first;It is more clinically significant to observe the aspect ratio of nodule A / T ≥ 1 in cross section.(2)The best prediction model of single slice of PTC ultrasonic image is: on the transverse plane of PTC,SVM classifier is used to simulate and analyze the data set,and then the 5-fold cross validation method is used to evaluate the effectiveness.When 17 LBP features are compressed,the classification prediction effect is the best and the accuracy can reach Acc= 0.9829.The results show that the transverse sonograms are more meaningful in PTC diagnosis.(3)The best prediction model of PTC based on ultrasonic image is: LBP and Hessian algorithm feature extraction and analysis are carried out from both of the longitudinal and transverse sonograms.The FS + C strategy is used to select the features.Finally,the accuracy of the model for the recognition of thyroid papillary carcinoma is Acc=0.9949.(4)Through the research on the above computer-aided diagnosis methodology,we try to make a preliminary design and implementation of PTC intelligent detection system.Verified by random samples,the program has good predictability for PTC sonogram and has certain auxiliary value for clinical diagnosis.Future research will continue to expand the range of diseases and further improve and optimize the detection system.The research innovations are as follows:(1)In this research,the application of artificial intelligence technology to improve the ultrasound image recognition technology,the study found that the transverse sonogram provides more information for PTC prediction diagnosis,which is consistent with the recommendations of ACR ti-rads grading guidelines,which is of great significance in the clinical diagnosis of thyroid nodules.(2)In this study,the optimal classification and prediction model of PTC ultrasonic image is established.We try to make a preliminary design and implementation of PTC intelligent detection system,and the prediction reliability is good.In the future,with the further combination of artificial intelligence and clinical diagnosis and treatment and the establishment of personalized medical database,the research will continue to optimize the prediction model and finally establish a mature tumor image AI detection technology,which has important scientific value for clinical diagnosis. |
PDF Full Text Request