基于改进遗传算法和LVQ神经网络包裹算法的特征筛选

Features Selection Based on MAGA/LVQ Wrapper Approach

在皮肤症状计算机辅助测试系统研究中,症状特征的筛选是提高系统诊断的关键问题,针对这个问题提出基于遗传算法和LVQ神经网络相结合的包裹算法。同时为了提高搜索效率,采用改进的自适应遗传算法。并用留一交叉法验证LVQ神经网络分类器的识别率,对初步提取的体现病态皮肤症状特点的22个特征以及它们的10个随意组合构成的干扰项进行特征选择,选择出使皮肤症状诊断率得到明显提高的特征组合。实验证明该方法是可行的。

In the research of computer-aided diagnostic （CAD） system for the skin symptom diagnosis, feature selection is a key process and affects the design and performance of the classifier. A new method based on Genetic Algo- rithms （GA） and LVQ Neural Networks wrapper approach is introduced to select rational feature from the originally 22 collected which has been extracted from the Skin Micro-image by image processing and 10 interferential items. At the same time, a modified adaptive genetic algorithm （MAGA） is introduced to enhance the searching efficiency. Furthermore, Leave-one-out cross-validation （LOOCV） scheme is employed to test the performance of LVQ Neural Networks classification, experimental results are satisfied.