基于专家乘积系统的组织病理图像分类算法

Histopathological Image Classification Algorithm Based on Product of Experts

组织病理图像的自动分类是医学图像处理领域的重要问题,有效特征提取方法是实现准确诊断的关键。为了实现组织病理图像的特征表示,提出一种基于专家乘积系统(PoE)的特征提取算法,利用最大似然和蒙特卡罗随机采样方法训练对应不同图像类别的Po E模型,将图像样本在所有模型下的响应相连作为其特征向量。根据训练图像样本的特征向量建立支持向量机分类模型。实验测试了宾夕法尼亚州立大学诊断实验室公开的组织病理图像数据库中的肾、肺和脾的健康及患病器官的组织病理图像,结果显示,所提算法在3种器官图像分类中均具有较高的准确性。

Automatic classification of histopathological image is vital in medical image processing field, and the effective feature extraction plays an key role to realize accurate diagnosis. A feature extraction algorithm based on Product of Experts （POE） is proposed to realize the feature representation of the histopathological image. The maximum likelihood and Monte Carlo random sampling methods are used to train PoE models corresponding to different kinds of images, and the responses of image samples in the two models are concatenated as their eigenvectors. Finally, a support vector machine （SVM） classification model is built based on the eigenvectors of the trained image samples. The experiments are carried out to classify histopathological images of healthy and inflammatory organs of kidney, lung, and spleen, which are provided by the Animal Diagnostics Lab at Pennsylvania State University. The experimental results show that the proposed algorithm can achieve high accuracy in three organ image classifications.