Research On Auxiliary Diagnosis Of Thyroid Disease Based On SPECT Image And Deep Learning

The incidence of thyroid disease in China is high and it is an important endocrine system disease.SPECT is a technique for imaging human function and metabolic information by means of a single photon nuclide labeled drug.Unlike ultrasound or CT,which diagnoses diseases through anatomical structures,SPECT can reflect the function of the thyroid gland and metabolic inf ormation in the body.Disease can be found earlier before the abnormality of the tissue structure,so SPECT has an irreplaceable advantage in clinical diagnosis.In most instances SPECT is the most important basis for the diagnosis of difficult diseases.In recent years,the rapid development of deep learning technology has become one of the important means of computer-aided diagnosis.Applying deep learning to computer-aided diagnosis of thyroid disease through SPECT image can effectively assist doctors to improve diagnostic accuracy,reduce work intensity and improve work efficiency.It is of great significance for the diagnosis and treatment of thyroid diseases.This paper works on the diagnosis of thyroid disease based on SPECT image and deep learning.Auxiliary diagnosis is performed on four thyroid diseases: hyperthyroidism,subacute thyroiditis,Hashimoto's thyroiditis,and subclinical hyperthyroidism by designing a convolutional neural network architecture and training algorithm and training the network model using the SPECT image dataset.The main contents are as follows:Firstly,the data augmentation and network architecture improvement of convolutional neural networks for the diagnosis of thyroid diseases are studied.The characteristics of SPECT image dataset are analyzed,and the data augmentation method is designed for the problem of less data volume.Based on the Dense Net network,the network architecture is improved,and the trainable weight parameters are added to the cross-layer connection path to improve the flexibility of network feature expression and extraction.The migration learning method is introduced.The network parameters are pre-trained using the Image Net large-scale data set,and then the network parameters are fine-tuned using the SPECT data set.The experimental results show that the improved Dense Net can extract the features of SPECT images better than the existing classification models on the SPECT dataset,and has better classification effect.Secondly,the improvement of the training algorithm of convolutional neural network is studied.The existing classical training algorithm is analyzed.A swarm intelligent optimization algorithm named as flower pollination is introduced in the stochastic gradient descent algorithm.The loss function on each batch data is used as the fitness function of the flower pollination algorithm.The hyperparameter in the training is used as the independent variable of the fitness function,and these hyperparameters are optimized by the flower pollination algorithm,so that the loss function value of each batch data in the training process is sufficiently reduced.The experimental results show that the algorithm can automatically adopt more suitable hyperparameters than the traditional stochastic gradient descent algorithm,so that the loss function can be further optimized,and the auxiliary diagnostic model can achieve higher precision.Finally,an incremental learning method of convolutional neural networks for the diagnosis of thyroid diseases is studied.This dissertation introduces the existing incremental learning method,and designs a dynamic convolutional neural network architecture based on the increasing of SPECT image data that can be acquired in the application process.It automatically searches according to the available SPECT image data.The optimal number of network layers makes the network architecture always adapt to the amount of data in the process of increasing data.The experimental results show that the dynamic architecture network can adjust the network layer number and fit the data properly on the data set of any size relative to the fixed architecture network so as to maintain the highest classification accuracy that can be achieved with current data without over-fitting or under-fitting.