Research On Multimodal Medical Image Recognition Of Thyroid Nodules Based On Deep Learning

Objective Using the convolutional neural network in the field of deep learning,a network model for judging benign and malignant thyroid nodules based on ultrasound images and CT images of thyroid nodules was designed,and the accuracy of current sonographers in judging benign and malignant thyroid nodules was compared,so as to analyze the depth of Whether the learning network model can help radiologists in judging benign and malignant thyroid nodules.Methods A total of 1082 ultrasound images and CT images of thyroid nodules from 922 patients in Zhejiang Cancer Hospital were collected,and the computer-identifiable regions of interest were marked by manually drawing inverse color processing of thyroid nodules in the images(Regions of Interest,ROI),and then divide the resulting data into training,validation,and test sets.Select a deep convolutional network model with the best performance from the five network models of ResNet34,InceptionV3,VGG19,DenseNet121 and AlexNet,use the training set to train a classification model with strong classification performance,and use the validation set to debug the classification model,and use the test set data to test the classification effect of the model.An additional five sonographers were selected to retrospectively diagnose the test set ultrasound images and classify thyroid nodules within the images into benign and malignant ones.The area under the operating characteristic curve(AUC),sensitivity,specificity,accuracy,positive predictive value,and negative predictive value of the test set were measured by the network model of the test set,the ultrasound image model alone,the ultrasound combined CT image model,and the five Diagnostic performance of sonographers for benign and malignant thyroid nodules.Results The depth convolution network model ResNet34 with good classification effect was screened out from the collected image data.the average accuracy,sensitivity,specificity and AUC value of the ResNet34 model using separate ultrasound images were 82.9%,0.836,0.820 and 0.889,respectively.The average accuracy of the ResNet34 model using ultrasound fusion CT image features on the test set is 85.0%,the sensitivity is 0.877,the specificity is 0.820,and the AUC value is 0.931.The average accuracy of the introduced ultrasound physician group is 74.4%,the sensitivity is 0.734,the specificity is 0.755,and the AUC value is 0.745.Conclusion In this paper,the effect of artificial intelligence in judging benign and malignant thyroid nodules is studied,and a deep convolution neural network model is trained with medical images,in which the model is effective in judging benign and malignant thyroid nodules in ultrasonic images of individual thyroid nodules.the average accuracy of the tested images is higher than the average level of the physician group.The model diagnosis effect of ultrasonic image fusion CT image feature is better than that of ultrasonic image only.This shows that the depth neural network model has higher performance,higher accuracy,sensitivity and specificity than a group of skilled ultrasound doctors in identifying thyroid cancer patients,and the training and recognition effect of cross-modal combined ultrasound and CT images is better than that of single ultrasound images.it is also said that the depth neural network model can assist imaging doctors in judging the benign and malignant thyroid nodules.It is meaningful to conduct clinical trials in the future.