BP-ANN在光学相干层析图像分类中的应用

为了研究反向传播人工神经网络(BP-ANN,back-propagation artificial neural network)对光学相干层析(OCT)图像的分类能力以及用不同算法训练的网络之间的性能差异,设计了基于纹理特征分析的BP-ANN图像分类实验系统。针对不同图像集,系统可根据类内和类间分散度的比值自适应地筛选最具区分性的纹理特征组成特征向量,再利用以不同算法训练的BP-ANN进行分类。实验表明,BP-ANN在经过快速训练后可以有效分辨不同组织图像,而Levenberg-Mar-quardt(LM)算法则被认为是最为有效的训练算法。以LM算法训练的BP-ANN可以在1 s内以平均8次的迭代计算完成训练,对测试集的分类准确率可以达到93.0%。

Application of BP-ANN in optical coherent tomography images classification

In order to confirm the classification ability of the back-propagation artificial neural network(BP-ANN) for the optical coherence tomography images and find the proper training algorithm for the BP-ANN,an image classification system based on texture features analysis is proposed.The texture features are firstly extracted from each image and then the most effective ones,which are automatically selected with the ratio of the within-class scatter to the between-class scatter,construct the feature vector for image classification.The experimental results show that the BP-ANN can be used to classify different tissue images and the Levenberg-Marquardt(LM) algorithm is thought to be the most effective training algorithm.The BP-ANN trained with LM algorithm can reach the convergence in just one second within 8 iterations,and can achieve the accuracy of 93.0% on average.