Research On Ultrasonic Diagnosis Of Benign And Malignant Thyroid Nodules Based On Deep Learning

Thyroid nodule is a common clinical disease and can be caused by various factors.And the key to cure it lies in the diagnosis of its being benign or malignant.As a common way to diagnose the thyroid nodule,ultrasonography plays a role in determining the size of thyroid nodules,locating the nodes before the puncture biopsy is administered.Further fine-needle biopsy or surgery should be performed when the radiologist determines that the thyroid is malignant based on thyroid ultrasound images.However,patients may incur unnecessary costs as inexperienced radiologists may make misdiagnoses.In addition,the variety and complexity of thyroid nodules,as well as the high noise and low contrast of ultrasonic images,also bring great challenges to doctors’ diagnosis.With the development of science and technology,computer technology has been applied to many fields and achieved good results.We know that computer processing images is accurate and efficient,which makes up for human errors.Therefore,in order to improve diagnostic accuracy,it is necessary to develop computer-aided diagnosis system.In this thesis,an in-depth study is conducted on thyroid nodule diagnosis based on Deep learning,aiming at the defects of existing Deep learning methods,relevant improvements were made.Moreover,despite good results appear in thyroid nodule analysis based on deep learning,in the view of clinicians,the diagnostic results are lack of interpretability because it does not use clinical information.Therefore,this thesis also tries to explore the interpretability of Deep learning methods in medicine.The main work is as follows:(1)The thesis,based on the characteristics of the ultrasonic image(which is taken at the variant position with multiple angles and high flexibility),put forward CNN(Convolutional Neural Network,CNN)which adopts the attention mechanism and retains the inter-pixel relationship maximumly.The CNN-SE-MPR model is integrated with the Squeezing-andExcitation module(SE)and the Maximum Retention of inter-Pixel Relations Module(MPR).Features can be adaptively selected from ultrasonic images and the relationship between pixels can be retained,thus ensuring the accuracy of the diagnosis.(2)In order to solve the problem of interpretability,we proposed a deep neural network model(MT-CNN)based on multi-task Learning(MTL).In this model,clinical ultrasonic features such as Contour,Margin,Calcification and Echogenicity were introduced into the network as four auxiliary tasks,and the benign and malignant judgment of thyroid nodules was set as the main task.The clinical features of the model were learned by auxiliary tasks and then transferred to the main task.After combining with the features of the main task,the final nodular classification was performed.