Recognition Of Benign And Malignant Thyroid Nodules Based On Improved DenseNet

It is of great significance for doctors to build an accurate diagnostic model of thyroid nodules with the help of ultrasonic images and deep learning technology while making efficient and accurate qualitative diagnosis.Aiming at the problems of insufficient training samples,difficulty in extracting deep feature of nodules and lower recognition rate of benign and malignant thyroid ultrasound images in the process of model construction,this thesis carries out the research on thyroid ultrasound image classification algorithm based on transfer learning and deep convolution neural network.The main research contents and phased results obtained are as follows.(1)Preprocess the thyroid ultrasound image data set,improve the image quality,and expand the number of samples,so as to solve the problems of poor original image quality and fewer samples.Firstly,mark the location of thyroid nodules in the ultrasound image by imaging experts,and then locate and extract the regions of interest of the nodules in batches to obtain the ultrasound image required in the follow-up experiment;Geometric enhancement methods such as random clipping,flipping and rotation are used to expand the data set.(2)Construct cascade deep feature extraction module,avoid the limitation of single model structure,improve the feature extraction ability and solve the problem of difficulty in extracting deep feature.Three kinds of pre-trained models(Alex Net,VGG16 and Res Net50)are obtained based on thyroid ultrasound image data set by using transfer learning and fine-tuning strategy,and then combined them into deep feature extraction module,so as to avoid the limitation of fixed scale input and network architecture of single model on deep image feature extraction.The experimental results show that the cascade feature extraction module realizes the complementary advantages of feature extraction of multiple models,and has better recognition performance than single model.(3)The depth separable convolution is introduced to improve Dense Net network,which is integrated with the cascade deep feature extraction module to further improve the classification performance and solve the problem of lower recognition rate of benign and malignant nodules.Replace the standard convolution of Dense Net network with depth separable convolution,and introduce Leaky Re Lu activation function to obtain Improved-Dense Net,which has lower complexity and faster convergence speed.Then,construct a complete thyroid ultrasound image classification model based on Improved-Dense Net and cascade deep feature extraction module,so as to further improve the classification performance of benign and malignant.(4)A comprehensive comparative experiment and result analysis were carried out,which based on the real collected thyroid ultrasound image data set.The comparative experiments show that the proposed model is superior to the traditional machine learning algorithm and the deep learning algorithm in other literature in accuracy,sensitivity and specificity.The recognition accuracy is 95.4%,and the sensitivity and specificity are 94.7% and 97.1% respectively.Experimental results show that proposed algorithm can identify benign and malignant thyroid nodules effectively.