The Research On Classification Of Ultrasound Breast Tumor Based On Semi-supervised Decision Tree And Fuzzy Inference

The ultrasonography is one of the major scheme for the early detection of breast tumor,and more and more scholars pay attention to the computer-aided diagnostic system based on the ultrasonography.However,the major ultrasound breast tumor classification method at present much relies on the low-level feature extracted from images.Their diagnosis process is hard to be understood by doctors.Due to the dependency of image quality,it is difficult for the previous method to deal with the ultrasound image collected from different ultrasound equipment.Meanwhile,the collection of labeled data is a tough task for the ultrasound breast tumor samples.Most of previous methods cannot hold the high performance facing the shortage of labeled data.These problems limit the application of the previous methods in clinical diagnosis.This paper introduces three novel ultrasound breast tumor classification methods to solve the problems of the previous method.Firstly,an ultrasound breast tumor classification method based on BI-RADS features scoring scheme and decision tree is suggested to build a diagnosis model with high comprehensibility for doctors.Then,a semi-supervised decision tree method based on support vector machine is proposed to reduce the dependency of labeled data.In the last,this paper introduces the ultrasound breast tumor classification method based bicluster and fuzzy inference.The bicluster algorithm is adopted to dig the diagnosis pattern,and the diagnosis rules are built based on these diagnosis pattern.The fuzzy inference is employed to integrate these diagnosis rules and give the final classification result.This method imitates the thinking of doctors during diagnosis,which is easier to be understood and gets a higher performance than previous method.Due to application of expert scoring scheme,three methods proposed in this paper are no longer limited by source of ultrasound image,and can deal with the ultrasound image from different ultrasound equipment.This paper utilized 1488 ultrasound breast tumor cases to evaluate three proposed methods.The experiment result shows that the proposed methods get higher performance in accuracy,sensitivity,and positive predictive value than previous methods,and can provide valueable diagnostic information for doctors.