基于K-L变换和神经网络分类的微钙化点检测新方法

A New Method on Detection of Microcalcifications Based on Principle Components Analysis and Neural Network Classifier

目前一般的乳腺X光片微钙化点检测系统大致都包括:图像预处理和分割;病理图像的特征提取和分类;辅助诊断和分析等几个步骤,其中神经网络经常用于特征提取和分类阶段。为了提高神经网络的分类能力,需要采用最具代表性的特征作为分类系统的输入部分,而且采用的特征数目要有利于最有效的特征提取,否则会使分类的效率大打折扣。所以分类系统一个重要的任务是对神经网络的输入样本集进行训练和特征值的优化,本文采用K-L变换用于降低输入特征向量的维数,从而达到参数优化的目的。试验表明,该方法可以有效地提高系统的灵敏度,降低诊断的假阳性。

At the present a regular mammographic microcalcification detection system is based on a three-step procedure： prepro- cessing and segmentation; feature extraction and classification; computer-aided detection and analysis. Typically, a neural network is usually used in feature extraction and classification procedure. In order to improve the classification ability of a neural network, we need accept the most representative features as input part. And the number of features must be helpful to the most effective feature extraction, or else the efficiency of classification will be greatly depressed. So an important ink of the classification module is to train the input data set of a neural network and optimize features. In order to optimize features, a PCA is applied to reduce the dimensionality of the input vector. The experiment results show that the method has extensively high assessments of their effectiveness in terms of sensitivity and reduction of false positive rate.