Research On Assisted Diagnosis Of Thyroid Nodules On SPECT Image Based On Deep Learning

In China,the incidence rate of thyroid diseases is high,but it has not attracted enough attention.Single-Photon Emission Computed Tomography(SPECT)is a functional imaging technology,which reflects the difference of activity intensity of human tissues by radioisotope drugs.It can detect diseases before the occurrence of abnormal human tissues,so it has its unique advantages in the actual clinical diagnosis.In recent years,the development of deep learning has greatly promoted the development of computer-aided diagnosis.Applying deep learning to thyroid SPECT images can reduce the work intensity of doctors,improve work efficiency,optimize the diagnosis and treatment process,save time and reduce costs,and learn the diagnosis and treatment experience of excellent professional doctors,To improve the accuracy of diagnosis is helpful to solve the problem of limited medical resources and fairness.In this paper,the auxiliary diagnosis of thyroid nodules based on deep learning SPECT images is studied,including model design,model interpretability method and software development.The main contents are as follows:Firstly,an auxiliary diagnosis model based on convolutional neural network is designed.The data collected from the hospital are cleaned to ensure consistency.According to the characteristics of less data sets and unbalanced data distribution,the appropriate augmentation method is selected.The improved Squeeze Net is adopted,and the channel attention mechanism is introduced on the basis of squeezenet to realize the importance screening of channel features.The Re LU activation function is replaced by the Mish activation function to ensure the information flow and make the gradient descent effect better.Experiments show that the recognition effect of the improved model is better than other mainstream networks.Secondly,the fusion model of convolutional neural network and manual feature is designed.The deep feature extracted by convolutional neural network is combined with the low-level and intermediate features extracted by hand,which can enrich the feature information.The texture feature and morphology feature are selected as handcraft features.The texture features are five descriptors based on gray level co-occurrence matrix,and the morphological feature are seven Hu moments.The fusion features obtained by combining manual features and depth features are sent into the classifier.The experimental results show that the feature fusion model has a better recognition effect on the classification of cool and other thyroid nodules(cold nodules and hot nodules as a class,namely other nodules).Good results can be achieved by using mainstream network classification in the classification of cold and hot thyroid nodules.Thirdly,the interpretability method of the model.The classification model of three kinds of thyroid diseases was explained by Gradient-weighted Class Activation Mapping(Grad-CAM).The experimental results show that the model pays more attention to which parts of thyroid diseases.This paper uses meaningful perturbation method to explain the two classification model of cold and hot thyroid nodules.Through the experimental results,we can see which pixels of cold and hot nodules affect the classification results of the model.Finally,Development of auxiliary diagnostic software.Medical image files are transmitted through DICOM protocol.The server(SCP)and client(SCU)are built through pynetdicom.After the SCP and SCU are built,the SCP files are transferred to SCU.QT is used to design the software interface and realize the basic functions of the software.The main functions are: visualization results of three classification and interpretable method of Graves disease,thyroiditis and thyroid nodules,visualization results of binary classification and interpretable method of cold and hot thyroid nodules,and binary classification of cool thyroid nodules and other nodules.