Research On Segmentation And Classification Of Thyroid Nodule In Ultrasound Images Based On Deep Learning

Thyroid nodule prevalence has risen over the years,and with the increase of people’s intake of iodine and the increase of radioactive pollution in the environment,the probability of its occurrence of cancer is also increasing,which has posed a serious threat to human daily health.During the examination of thyroid nodules,doctors often use ultrasound images to detect them.In this thesis,artificial intelligence is used to accurately segment nodules in thyroid ultrasound images and distinguish benign and malignant nodules,which can help doctors improve the efficiency of diagnosis.For accurate segmentation of nodules in ultrasound images,this thesis uses the Mask R-CNN network to segment the nodule,and optimizes the backbone network to strengthen its feature extraction function.An attention module is introduced into the backbone network,and the detailed insertion position is the residual part,which can speed up the convergence of the network,and add a branch to the feature pyramid model of the original backbone network,and then fuse all the output features.image,to obtain joint features,and finally output,so as to achieve multi-scale feature fusion and balance the information differences in each output feature map.Using 6840 ultrasound images of thyroid nodules to test the improved network model,the final model results,the average segmentation Dice coefficient and precision are 0.9562 and 0.9601,the average segmentation recall rate and F1 score are 0.9437 and 0.9518.In terms of the average division Dice coefficient,the improved network is 9.13% higher than the original network.The results show that the improved Mask R-CNN in this thesis has a high segmentation accuracy,which can be used in the clinical diagnosis of thyroid nodular diseases,and can improve the diagnosis efficiency of doctors.In order to achieve accurate classification of benign and malignant thyroid nodules,this thesis uses Dense Net network to detect ultrasound images and optimize its network structure.Five branches are added to the original network to fuse the input and output feature maps to transmit a large amount of underlying texture information.An attention mechanism is added between adjacent transition layers to completely retain the effective features of the image,so as to balance the information proportion after fusion of multi-scale feature maps,enabling the network to better utilize the global information of feature maps.Tested on 4000 ultrasound images of thyroid nodules,the average classification and recognition accuracy of the improved Dense Net is 0.9671,the average precision is 0.9684,the average recall rate and F1 score are 0.9711 and 0.9697,respectively.As far as the recognition accuracy is concerned,the improved Dense Net is3.77% higher than its original network.The results show that the improved Dense Net can predict benign and malignant thyroid nodules and shorten the time for doctors to check.