Research On Assisted Diagnosis Of Thyroid Nodules Based On Attention Mechanism And Feature Fusion

In recent years,the incidence of thyroid nodules has increased significantly,among which malignant thyroid nodules cause great harm to the human body.Ultrasonography provides an important imaging basis for diagnosing thyroid nodular lesions,but the diagnosis results rely entirely on the clinical knowledge and subjective judgment of doctors,which may lead to misdiagnosis.With the rapid development of artificial intelligence,using deep learning to identify the benign and malignant nature of thyroid nodules and assist physicians in decision-making has important clinical value.In order to improve the accuracy of diagnosis of benign and malignant thyroid nodules,the thesis uses the Efficient Channel Attention module and feature pyramid networks to improve the Res Net50 network.A neural network model with excellent recognition effect is obtained by training the model.Finally the improved model is applied to thyroid ultrasound image assisted diagnosis system.The specific research content of this article is as follows:(1)Under the premise of protecting the privacy of patients,a total of 1550 ultrasound images of real thyroid nodules in third-class hospital were collected.The important areas of thyroid nodules were marked with the assistance of professional doctors in the thyroid field and distinguished between benign and malignant nodules.This thesis selects three convolutional neural network models that perform well in the field of image classification to research.(2)In order to solve the problem of over-fitting that may occur due to less dataset samples,this thesis combines the transfer learning method,selects the pre-training model on Image Net and fine-tunes the model on the dataset.The experimental results show that the combination of transfer learning can significantly improve the recognition effect of the model on benign and malignant thyroid nodules,and avoid overfitting to a certain extent.Through comparative experiments,Res Net50 with the best recognition effect is selected as the backbone network.(3)In order to further improve the classification effect of the model on benign and malignant thyroid nodules,an efficient channel attention module is embedded in the Res Net50 network.It makes the model to extract more efficient features from the image.Aiming at the problem that the detail features of the thyroid nodule ultrasound image are seriously lost as the number of network layers deepens,this thesis uses a feature pyramid structure to fuses the features with different resolutions extracted by the network.Experiments show that the accuracy of identifying the benign and malignant thyroid nodules in the optimized model reaches 94.56%,which is 4 percent higher than before optimization.(4)The thesis applies the improved model to the thyroid ultrasound image assisted diagnosis system.In order to assist physicians in diagnosis,improve medical standards and facilitate patients,the thesis applies the improved model to the the system based on the software engineering method.Finally,using black-box method to test system.Use cases are written to test the system functionally.The test results show that the system functions as expected and can be applied to clinical practice.