Research On Auxiliary Diagnosis Technology Of Thyroid Nodules Based On Deep Learning

Thyroid nodule refers to a mass in the thyroid gland which has become a common disease in clinical medicine.There are a great number of inducements can cause thyroid nodule disease of which the cancer incidence is on the rise period according to lots of researches.The clinical diagnosis of this disease usually schedule most of nodules through ultrasound imaging.When detecting suspicious nodules,fine needle would be used for further biopsy.However,due to the mechanism of ultrasound imaging this method would produce strong image noise,ultrasound images have the characteristics of blurred borders and complex backgrounds.Meanwhile,biopsy is intrusive and accompanied by uncertainties.Therefore the traditional method of diagnosis would lead to not only misdiagnosis caused by the subjective judgements from doctors but also unnecessary biopsy surgery,which undoubtedly brings greater stress and anxiety to patients,increases additional medical expenses and manpower as well.In this thesis,we studied an auxiliary diagnosis technology of thyroid nodules based on deep learning which includes the segmentation of medical images,the classification of thyroid nodules and the analysis of the characteristics obtained from feature maps of the network.This method aims to achieve the accurate segmentation and classification of thyroid nodules via an end-to-end approach.The main contents and results of this thesis are as follows:(1)Researching on deep learning models with respect to the segmentation of medical images and designing the network applied to the segmentation of thyroid nodules images.The network proposed in this thesis fi rstly obtains feature maps of various sizes after down-sampling operation including convolution and pooling.Then this model introduces Self-Attention mechanism of Transformer model at the low layer of the structure to encode the feature map.The optimized U-Net makes up for deficiency the underuse of pixel space dimension information existing in traditional U-Net by the application of encoder,moreover it adopts the method of image restoration through up-sampling stitching from U-shaped structure network.In addition,we compared the convergence and segmentation accuracy of different loss function and proposed Dice-Cross Entropy loss function to improve the training of the model.(2)Researching and designing the deep learning models regarding the classification of thyroid nodules.In this thesis,we set up experiments to compare the classification accuracy and AUC of several networks designed for image classification such as VGG-16,ResNet-50 and the transfer learning of them.Later the transfer learning of ResNet-50 is regarded as the base model of the classification task.Then we made use of the concept of ensemble learning.We toke the original medical image of thyroid nodules and the feature maps from segmentation model as input images to obtain three feature vectors respectively which would be merged by fully connected layer to get the classification results.(3)Researching and analysis on feature visualization of thyroid nodules.As for the feature extraction in the network of classification of thyroid nodules we adopt forward propagation to get feature maps of every layer.The texture and edge features from shallow layer are helpful for inspecting some important features of thyroid nodules such as edges,cystic-solid and calcification.Moreover,we also used back propagation to visualize the convolution filters so that we could understand different feature types of different filters which increase the interpretability and reliability of the model.