Research On Artificial Intelligence And Multimodal Imaging In Some Disease Analysis

With the development of society,health has increasingly become a problem which people concern.In the diagnosis and analysis of diseases,especially tumors and neurodegenerative diseases,the following problems exist:doctors need to make a diagnosis by comprehensively diagnosing the patient's condition according to multi-modal medical images and clinical characterization which takes a lot of time and need experience accumulation;while computer aided diagnosis is mostly directed to a single modal image,ignoring other diagnostic reference data.This dissertation uses artificial intelligence and radiomics to solve the problems of current diagnosis.Through the development in recent years,machine learning and deep learning in artificial intelligence have become increasingly mature,and the application in the medical field has also had achievements.Similarly,since radiomics was proposed,it has been widely used in the diagnosis and research of diseases.Therefore,this dissertation selects multi-modal image data and clinical data and combines machine learning or deep learning methods to make diagnosis of breast cancer,Parkinson's disease and Alzheimer's disease.The specific work is as follows:1.This dissertation applies machine learning method to classify the benign and malignant masses of mammography images,establishes six machine learning models,and classifies the BCDR-F03 data sets separately.Then assess the experimental results of six supervised machine learning classifiers detailed.The experiment found that the optimal combination of training sample size and feature dimension can reduce feature dimension and significantly improve the accuracy of results.Through Repeatability studies,random forests have the highest classification accuracy during the test.Artificial neural networks and support vector machines have better generalization ability and stability.2.This dissertation refers to LeNet-5 design CNN for PD diagnosis,then trains CNN based on T2-MRI single-mode image set,PET single-mode image set and multi-modal fusion image set,and uses cross-validation method to obtain average diagnostic accuracy.Then change the size of CNN convolution kernels,the number of convolution kernels,the number of layers of neural network,replace the excitation function that can learn autonomously,add the dropout layer,use 1*1 convolution kernel and increase the pre-training process for transfer learning.These methods improve the network to get better performance and more generalized.The results of six experiments of CNN before and after improvement show that the diagnostic accuracy of CNN for multi-modal fusion image set is higher than that of single mode image set.The improved CNN diagnosis is much better than original CNN.So improved CNN is more suitable for diagnosis of PD.Finally,the feature map which is extracted by CNN,is visualized to assist doctors in diagnosis.3.This dissertation selects three kinds of CNNs,AlexNet,GoogLeNet and U-Net,and improves them.Computer-aided diagnosis of AD is performed based on heterogeneous data,and select the CNN which has optimal performance.Heterogeneous data includes multi-modal image data and clinical scale data.For AlexNet,increase the ability of extracting features by increasing the network depth and changing the number of convolution kernels;for GoogLeNet,the feature channels are weighted by using the Squeeze-and-Excitation module;for U-Net which is specifically for medical image segmentation,extract features of the segmented image which is the output of U-Net,and then use SVM to complete classification.The training of the three networks adopts transfer learning,and the initial parameter values are obtained through pre-training to realize deep learning based on small sample medical images.Use 10-fold cross-validation method to compare the computer-aided diagnosis results.The results of T2-MRI and PET single-mode images were compared with the results of using SVM to classify based on T2-MRI,PET features which are extracted by CNN,and clinical scale data.The results show that the improved CNN classification performance based on U-Net is the best,and the diagnosis based on heterogeneous data is better than the diagnosis based on single modal data.This dissertation makes full use of the method of radiomics,multi-modal images and heterogeneous data to obtain more comprehensive patient information,using machine learning and deep learning methods to apply artificial intelligence in the medical field,assisting doctors to make diagnosis,and provide better treatment options to improve the quality of life and survival of patients.Through artificial intelligence to achieve computer-aided diagnosis and make it popular are conducive to the full utilization of medical resources,and also help to form a harmonious relationship between doctors and patients and promote the harmonious development of China's health.