Research On Auxiliary Diagnosis Of Thyroid Nodules Based On Ultrasound Images

Thyroid nodule is one of the most common nodular diseases,the incidence of which is increasing all over the world.At the same time,it shows a trend of increasing year by year in our country.The harm of malignant thyroid nodules is extremely serious,but Early diagnosis of malignant thyroid nodules can greatly reduce the potential risk.In the past two decades,the detection technology of thyroid nodules has developed rapidly.Highresolution ultrasonography has become an indispensable method for early screening and diagnosis of thyroid nodule due to its advantages of non-surgical,rapid,accurate and convenient.However,the traditional ultrasonography focuses on experience-based imaging features,which are cumbersome,subjective and highly dependent on the experienced clinical radiologists.In the meantime,ultrasound images are noisy and lowcontrast,and different equipments will also cause differences.These problems will affect the diagnosis results and even lead to misdiagnosis,which bring huge challenges to doctors.With the continuous development of deep learning technology and the huge improvement in image classification,it has brought new ideas for diagnosis of benign and malignant thyroid nodules.In this thesis,deep learning methods are adopted to study the segmentation of thyroid nodule lesions,multi-type feature extraction and classification based on clinical ultrasound images of real patients.Besides,this thesis also introduces transfer learning,ensemble learning and attention mechanism.In the aspect of nodule lesion segmentation,this thesis adopts U-Net as the basic network and conducts a comparative experiment with its variant U-Net++,and modifies its structure in some respects.Under the condition of keeping the symmetric structure unchanged,the Encoder is modified and adds residual structure,and this thsis names it Res U-Net.The results show that,adopting Res Net as backbone improves the accuracy of segmentation to a certain extent.In the aspect of nodule feature extraction and benign and malignant classification,in order to cope with the problem of insufficient medical image datas,this thesis innovatively adopts the transfer learning and ensemble learning to build an ensemble model based on Alex Net,VGG16 and Res Net50.Under the Stacking strategy,this model can extract the shape features,texture features and global features which are fused to obtain higher accuracy than the single feature and model.In order to get a better accuracy in thyroid nodule diagnosis,this thesis introduces the Convolutional Block Attention Module which adds channel attention and spatial attention to some traditional convolutional neural networks.In the meantime,this thesis ensembles the attention models again.Finally,this thesis verifies its influence and performance on different models.After the introduction of the attention mechanism,the accuracy of the model has been improved to a certain extent,which finally reached84.28%.