摘要：Aims: The main purpose of this research was to develop a new method to extract the most valuable texture features to differentiate between cyst and solid nodule classes in breast echography images. T -test coupled with leave-one-out cross-validation analysis technique was developed. This technique was used to a breast ultrasound image database in order to select a small number of highly predictive features and to allow algorithms to operate effectively and faster. Study Design: The image processing was made using the Matlab environment and statistical analysis was accomplished by using the SPSS ver. 17 software. Place and Duration of Study: Department of Echography, St Maria’s Hospital, Galati, between November and December 2011. Methodology: To reach this goal, a feature extraction method was developed based on the geometric and statistical moments. Features extraction has been successfully accomplished and their further application has been based on an integrated statistical approach. To determine the meaningful features and their efficiency in each studied class, the statistical T -test was carried out. T -score was performed to establish the capability of the features to differentiate between classes and also, as a ranking tool for features. In order to analyze if our results would lead to an independent data set, the leave-one-out cross-validation method has been used. Results: Experimental results showed that the proposed method is very effective and the selected feature subsets could be used to compare the ability to differentiate between classes. Also the minimum size of the feature subsets was another pursued goal. Three different combinations of the statistical and geometric moment features could characterize the breast nodule (i.e. rectangularity, area convexity and the second order moment) and other three could characterize the breast cyst (i.e. circularity, form factor and eccentricity). Conclusion: Through this method, the dimensionality of the feature vectors was substantial reduced and the ability to differentiate between cyst and solid nodule classes in breast echography images was improved.
