| Objective:To collect the titles from the RIS system of the University City Hospital Affiliated to Chongqing Medical University and the PACS system of the First Hospital Affiliated to Chongqing Medical University from the deputy chief physician from October 2017 to January 2020.The ACR TI-RADS classification has been given Thyroid ultrasound image,using the convolutional neural network in the field of deep learning to design an ACR TI-RADS diagnostic model based on thyroid ultrasound image,to provide clinical ultrasound doctors with auxiliary diagnostic reference,can also be used for interns and young veterinary ultrasound doctors Teaching aids.Method:Collect 150,000 thyroid ultrasound images of 2000 cases of thyroid ultrasound images with a title of the deputy chief physician from October2017 to January 2020,which have been given ACR TI-RADS classification,and undergo image preprocessing and data Enhanced data set processing methods to obtain a data set that can be used for convolutional neural network model learning,using transfer learning strategies,Cosine Learning Rate Decay(Cosine Learning Rate Decay),model fusion and other methods to design and train a strong classification The performance classification model and the evaluation model’s classification effect are tested using the test set data.At the same time,two groups of primary and intermediate doctors are introduced to read the test set,and the accuracy and efficiency of the three groups are read.Results:After the screening of images,a total of 4300 thyroid ultrasound images from 2000 cases were finally collected in this experiment.ACR TI-RADS classification was 300 in Class 1,500 in Class 2,500 in Class 3,and 500 in Class 4.,300 pictures in 5 categories,of which all 5 cases collected were confirmed as malignant by pathological results.After data balance processing and data enhancement,the final data volume is 36000.Use the data set to train and screen the five convolutional neural network models that have been pre-trained using transfer learning methods,VGG,Inception,Res Net,Dense Net121,and Xception.The model Dense Net121 with the best classification effect is used,and then the Cosine Learning Rate Decay and model fusion methods are used to make the highest classification accuracy rate of the model reach 95.83%,the lowest accuracy rate is83.95%,and the average accuracy rate is 87.87%.After reading the pictures,the statistical analysis of the doctor reading group showed that the overall accuracy of the model was significantly higher than that of the junior doctor reading group,which was the same as that of the intermediate doctor reading group.Conclusion:This topic takes the application of artificial intelligence(AI)technology in medical image recognition as the research background.Based on thyroid ultrasound images and using deep learning methods,the intelligent diagnosis model designed and trained has a good effect on the classification of thyroid nodules.And high efficiency and good stability.However,there are also deficiencies in the research of this subject.The collected data set is not large in scale and wide in scope,so there are limitations in reflecting the classification performance of the model.In the future research,the ultrasound department of more hospitals can be combined to increase the size and size of the data set,so that the model is more clinically practical and can be used as a reference for clinical ultrasound doctors. |
PDF Full Text Request