Diagnostic Research Of Ultrasound Thyroid Nodules Based On Deep Learning

Thyroid disease is one of the most commonly diagnosed nodular lesions in the adult population.The incidence rate is increasing year by year.How to accurately diagnose thyroid nodules has attracted a lot of attention around the world.Because ultrasound examination has the advantages of real-time detection,low cost,no damage to the human body,and repeatable detection process,it is the most commonly used imaging method for diagnosing thyroid diseases in hospitals.However,the contrast of ultrasound images collected by the current machine is very low,and the morphology of thyroid nodules is complex and diverse,which brings great challenges to the diagnosis of doctors,and computer-aided diagnosis technology is urgently needed.This thesis uses deep learning to improve the related techniques of computer-aided diagnosis of ultrasound thyroid nodules.The main research content consists of three parts:segmentation of ultrasound thyroid nodules,feature extraction of nodule regions,and classification of nodules.Due to the low contrast of medical ultrasound images and the complex and diverse thyroid nodules,the traditional image segmentation algorithm is not ideal.This thesis proposes a marker-guided deep network segmentation model.For the nodule region,this thesis proposes a method based on deep learning for shape and texture feature extraction.Finally,the feature selection is made for the merged features,and the nodules are classified by logistic regression and support vector machine.The main contributions of this thesis are as follows:(1)This thesis proposes a marker-guided segmentation method,which combines the marker information in the daily diagnosis of doctors to segment the thyroid nodules based on a full convolutional neural network,and the segmentation accuracy has been significantly improved.(2)This thesis proposes a method for extracting the shape feature of thyroid nodules based on a convolutional autoencoder,and obtains better classification accuracy than traditional shape features.(3)This thesis proposes a texture feature extraction method based on feature maps of deep network and bag of words model,which is superior than traditional texture features in identifying nodules.(4)In this thesis,feature selection is made for the shape feature,texture feature and echo feature.Logistic regression and support vector machine are used to classify the nodule.The highest average accuracy is 77.2%.