| Medical image diagnosis is an important part of clinical diagnosis.At present,clinicians mainly detect the lesions by observing the two-dimensional slice sequence medical images of patients.This traditional diagnosis method seriously depends on the level of doctors’ experience,and has the disadvantage of strong subjectivity,low accuracy and high misdiagnosis rate.In recent years,artificial intelligence technology has developed rapidly,comprehensively innovated algorithms in the field of image and video,and has played an important role in industries,such as intelligent transportation and smart city.To face the needs of national precision medicine strategy,artificial intelligence-aided medical image diagnosis has become a research hotspot.With the advantages of big data and deep learning algorithm in medical image,artificial intelligence technology can greatly improve the accuracy of clinical diagnosis and reduce the probability of misdiagnosis and missed diagnosis.However,at present,most artificial intelligence-based diagnosis methods in medical image are directly transferred from the field of computer vision,which cannot fully mine the potential information for clinical diagnosis in medical image data,and ignore the diversified basis of clinical data and the requirements of decision transparency.Firstly,in addition to the tumor itself,the surrounding environment of the tumor will also have an impact on the occurrence and development of the disease.Exploring the diagnostic value of the peritumoral environment and innovatively building the diagnostic model of fusion peritumoral images is a new direction of medical image diagnosis;Secondly,the clinical data used for disease diagnosis is not only limited to medical images,but also includes diversified data,such as blood test indicators,sign information and risk factors.How to efficiently integrate multi-source heterogeneous data for individual patients to improve the diagnostic accuracy is a difficult problem to be solved at present;Finally,a disease or lesion often presents many signs in imaging.It is very important to evaluate a variety of related signs simultaneously,explore the relationship between disease and signs,and improve the accuracy of disease diagnosis.Aiming at the above three problems,based on artificial intelligence technology,this dissertation takes the prediction of lymph node(LN)metastasis of T1 lung adenocarcinoma(LUAD)as an example to carry out relevant research.The main work and contributions include:1.Aiming at ignoring the surrounding environment of tumor in current clinical diagnosis,this paper designs experiments and verifies the important value of peritumoral region characteristics in clinical diagnosis.851 imaging features were extracted from the tumor region and peritumor region respectively.The feature dimension was reduced by correlation analysis.The model of LN metastasis prediction of T1 LUAD was established by least absolute shrinkage and selection operator(LASSO)and Support Vector Machine(SVM).The results showed that both tumor and peritumor features can be used to predict LN metastasis for T1 LUAD,and the classification ability of the model based on peritumoral features is equivalent to that based on tumor features.In addition,this study fused tumor and peritumor information through three different information fusion strategies,image level,feature level and decision level,respectively,to explore the optimal information fusion method for this study.It can be found that the method based on image level information fusion has the best performance,with AUC values of 0.923 and 0.762 in internal and external test datasets,respectively.This study verified the important value of peritumoral region,and provided an important reference for clinical decision-making and personalized treatment plan.2.Facing the diversified medical data such as clinical images,biochemical blood test indexes,clinical characteristics and risk factors,this paper proposes a neural network model of multi-source heterogeneous data fusion to improve the accuracy of clinical diagnosis.Through the end-to-end architecture design,the AUC values of the network framework designed in this study in internal and external tests reached 0.946 and 0.857 respectively,which significantly improved the performance of the model compared with doctors and imaging methods.In addition,with the help of deep model visualization method,this study quantifies the attention of the model to diversified data,improves the interpretability of the model,and effectively guides and assists doctors in the diagnosis of LN metastasis.3.Aiming to deeply explore the correlation between diseases and signs,and improve the diagnostic performance of the model.This paper proposed a multi-scale,multi-task,and multi-label classification network(3M-CN).The model accurately predicts LN metastasis for T1 stage LUAD of patients,as well as evaluate multiple related signs of pulmonary nodules,including lobulation sign,spiculation sign,pleural indentation,and attenuation.3D densenet was used as the backbone of the model to extract the three-dimensional features of pulmonary nodules.With the help of the multiscale feature fusion(MFF)module,the model can adapt to the personalized dependence of multiple labels on different levels of image;With the help of the segmentation module,the model is guided to focus on the key areas for diagnosis;The proposed refine layer(RL)module can effectively fuse the risk factors into the neural network to improve the performance of the diagnosis.The results showed that the proposed three modules can effectively improve the prediction accuracy of LN metastasis,and the proposed 3M-CN integrating the three modules has the highest performance in the internal and external test datasets,with AUC values of 0.945 and 0.964,respectively.In addition,the model can provide more relevant semantic interpretation,which greatly enhance the interpretability of the deep learning model,increase doctors’ confidence in the model results,conform to doctors’ diagnostic process.It is of great clinical application value. |
PDF Full Text Request