Feature Extraction And Selection In CT Image Based Hepatic Lesion Diagnosis System

The proposed liver lesion diagnosis system mainly consists of 4 steps: (Rejion of Interest) ROI extraction, feature extraction, features selection and classification modules. Multi-phase abdominal CT images are first fed into a image registration module to eliminate their different in spatial positions. Then experienced radiologist could draw the ROIs just using one of the images. After that, all the features are extracted based on the ROIs. Feature selection module is then applied on some of the features to create the feature set. Finally (Support Vector Machine) SVM-based classifier is used to categorize the type of lesions.A huge difference between Liver (Computer Aided Diagnosis) CAD and other CAD like breast CAD and lung CAD is that shape prior has no effect in the detection and diagnosis of hepatic lesions due to the fact that liver disease is prone to diffusing–like and even the same type of the disease always varying greatly in shape from case to case. Therefore texture-based features become the optimal candidate target for feature extraction task. In this paper, Image texture feature extraction method includes first-order statistics (FOS), Gray Level Co-occurrence Matrix (GLCM) and temporal method. For image recognition, more inputted characteristic items do not means better, they may produce a lot of false positive findings."Information overload"will weaken the classification performance. In addition, when inputted features are increased, the training samples required for classification will grow in exponential. Therefore, in the liver diagnose system, how to choose the right feature set which contribute to the high classification accuracy from a number of features is the key issue.The basic task of feature selection is how to find out the most effective feature subset from high-dimensional features. It includes the following two sub-problems: Search strategy and the issue of evaluation functions. Genetic Algorithm (GA) is the randomized algorithm; the method is a searching method for solving in Local minima which adds randomly. (3) Sequential algorithms are added or subtracted features sequentially, Such as Sequential Forward Selection (SFS), Sequential Backward Elimination (SBE), Plus l Take-Away r Selection (PTA), Sequential Floating Forward Selection (SFFS) and Sequential Floating Backward Elimination (SFBE) etc. On the other hand, there are two types of feature selection framework, derived from the nature of the evaluation function J ? used: filters and wrappers. In a filter framework, J(?) measures the performance of a feature set in a manner that does not include the classification algorithm which will eventually use the features. In a wrapper framework, J(?) incorporates the classification algorithm.In this paper, based on filter and wrapper method, Support Vector Machine (SVM) and Genetic Algorithm (GA), a new method that utilizes the two-step selection approach was proposed to choose the most relevant features from a large feature set. This two-step selection method can be described as: firstly, apply traditional sequential algorithms such as SFS, SBE, SFFS, SFBE, PTA to obtain five different feature subsets which will be used to generate a new feature set and then utilizes GA to search feature space from the new feature group by the fitness function designed by the accuracy of SVM. The advantages of this approach include the ability to accommodate different feature selection search strategies and combine filter and wrapper method, which makes the system can find a small optimal feature subsets that perform well for a particular inductive learning algorithm of interest to build the classifier.The main innovation points are as follows:1. The use of multi-phase liver CT images.2. The use of a variety of assessment criteria.3. Accommodate different feature selection search strategies and combine filter and wrapper method.4. The use of support vector machine and genetic algorithm.