| Thyroid cancer is a disease with nodules in the thyroid region of the neck as the first symptom.It is one of the cancers with the highest incidence rate.Therefore,timely detection of thyroid nodules and identification of benign and malignant nodules can help prevent and operate thyroid cancer.With the rise of computer-aided diagnosis technology,the use of deep learning technology to automatically extract features and analyze thyroid medical images can help doctors to better analyze the condition and reduce subjective errors.At present,the mainstream clinical detection methods of thyroid nodules are mainly divided into ultrasound detection and CT examination.However,their images have their own defects,making the identification and classification of thyroid nodules still very challenging.In this paper,the characteristics of thyroid ultrasound images and CT images are studied and carried out respectively:(1)The neural network classification model of ultrasonic image based on residual network is studied,and an improved Res Net50 model is designed,which uses the deep network to extract the overall comprehensive features of the image.Aiming at the problem of blurred nodule boundary and salt and pepper noise in the image,adaptive histogram equalization and median filter are designed for data preprocessing.And by designing an experimental group,comparing Le Net、VGG16、Res Net and other networks,the accuracy rate of 82.54% in the open source dataset TN-SCUI2020 is achieved,and its performance is higher than that of each comparison network,which has certain practical value.(2)The deep learning-based image semantic segmentation network is studied,and a convolutional neural network based on encoder-decoder structure for CT image nodule segmentation is designed.The network structure uses the U-Net network model as the basis and combines Efficent Net as the backbone network,which can accurately distinguish thyroid nodules and background regions,and achieve high-performance segmentation.A new hybrid loss function is used,which combines the Dice loss function and the weighted cross-entropy loss function to accelerate the convergence process of network training and optimize the segmentation accuracy to a certain extent.(3)The CT image nodule classification network based on fusion network is studied,and combined with the segmentation network,the automatic diagnosis system of thyroid CT image is realized.This paper uses the bilinear pooling technology of network fusion to independently design a low-level and high-level feature fusion classification network,which can comprehensively learn the performance of thyroid nodules at different feature levels and achieve accurate classification of benign and malignant.And a smooth concatenated image processing algorithm is designed to meet the adaptive adjustment of the segmentation network output image to the input of the classification network,realize the intelligent segmentation of multi-nodular CT images,and achieve end-to-end diagnostic classification without any user intervention on the image.The data set collected by a hospital in Jiaxing City was used for experimental verification,and compared the results of the hospital’s professional diagnosis and judgment,the accuracy of the identification and classification system reached 85.92%. |
