Convolutional Neural Network And Transfer Learning In Medical Image Analysis

Based on proper design and training with sufficient medical instances,artificial intelligence assisted diagnosis systems enable accurate classification of benign and malignant lesions seen in images.That can reduce doctors' daily workload,decrease unnecessary biopsies and enhance the quality of patients' life.Furthermore,artificial intelligence assisted diagnosis systems are capable of cancer staging that facilitates the treatment planning and delivery and accordingly,the survival rate increases.Moreover,based on retrospective study of cancer cases,artificial intelligence assisted diagnosis systems are able to estimate the survival time.Thereby,precise medicine is possible and the medical resource can be made full use of.In the field of artificial intelligence,deep learning has recently revolutionized the image representation that has great potential applications in the field of medical imaging.Thus,in addition to the basis of traditional machine learning and feature engineering,this paper adds convolutional neural networks and transfer learning to medical image analysis.Specifically,this paper focuses on artificial intelligence assisted diagnosis systems and we find that,1.Transferred high-level features improve the performance of lesion diagnosis.Deep convolutional neural network is capable of strong abstraction of features and shows good performance of object classification.It necessitates large-scale instances for hyper-parameter optimization,while the collection of sufficient high-quality medical instances is infeasible.To improve the performance on limited medical instances,pre-trained deep convolutional neural networks were fine-tuned.The experimental results indicate enhanced performance,while a lot of time and effort is required for knowledge transfer.2.Feature selection benefits lesion classification.Massive features for image representation enhance the prediction but easily cause over-fitting,while fewer features suggest good generalization of artificial intelligence assisted diagnosis systems.For feature selection,Wilcoxon rank sum test is employed for feature ranking.In comparison with machine learning classifiers,artificial neural network robustly increases the classification result with a few features.Besides,more instances for training,better diagnosis performance.3.A combination of support vector machine and texture analysis can assist the differentiation of cancer subtypes.Based on different magnetic resonance imaging sequences,four kinds of texture features(gray level co-occurrence matrix,GLCM;gray level size zone matrix,GLSZM;gray level run length matrix,GLRLM;and multiple gray level size zone matrix,MGLSZM)and the classifier of support vector machine are inverstigated.Experimental results indicates that the texture analysis is helpful in clinical decision making,while the diagnosis performance can be further improved.It can improve the life quality,survival rate and time.In addition,it can optimize the existing medical resources by making full use of existing medical imaging data and unearthing the potentiality of artificial intelligence techniques in medical community.When the clinical decision-making is faciliated and patients are provided with accurate treatment plans,the doctor-patient relationship will be harmonized and the national medical and health service will be pushed forward.