Research On Thrombosis Recongnition Method Based On CAD And Cardiac Ultrasound Image Sequences

Cardiac thrombus is a serious threat to human health. With the rapid developmentof computer-aided diagnosis technology, combining cardiac ultrasound image sequencesand computer-aided diagnosis technology to detect thrombosis becomes an inevitabletrend. Pattern Recognition Research Center and the First Affiliated Hospital of HarbinMedical University have been researching on the identification and classification ofthrombosis and comb muscle since2008. On this basis and with the introduction ofanother two classes, which are normal and SEC, this paper researches on theclassification and recognition of four categories of cardiac ultrasound image. In thispaper, optical flow calculation method is used to transform the four-class-classificationproblem into two binary classification problems, one is the classification of thrombosisand pectinate muscles, the other is the classification of normal and spontaneousdevelopment. This paper presents a method of making a displacement matrix based onoptical flow and three features based on the displacement matrix. It also presents asimple and effective classification strategy, so that the classification accuracy reaches93.27%. The next step is classifying two binary classification problems. In order to dothat, this paper started from the optimization of feature extraction and the improvementof classification methods simultaneously. First, this paper uses some GLCM-basedtexture features and extended local binary pattern operators texture features forclassification. It achieves a high accuracy when use support vector machines andartificial neural network classifiers to classify normal and SEC. Taking into account thatthe sparse representation method is not sensitive for feature extraction method, so thispaper uses the classification method based on sparse representation to classifythrombosis and comb muscle. This paper proposes two non-negative constraints andtwo classification strategies, so that the classification accuracy of classifying thrombosisand comb muscle reaches91.93%, significantly higher than other classifiers.