图正则化字典对学习的轻度认知功能障碍预测

PDF下载HTML

Title:

Dictionary pair learning with graph regularization for mild cognitive impairment prediction

作者:

魏彩锋^1,2, 孙永聪^1,2, 曾宪华^1,2

1. 重庆邮电大学 计算机科学与技术学院, 重庆 400065;
2. 重庆邮电大学 计算智能重庆市重点实验室, 重庆 400065

Author(s):

WEI Caifeng^1,2, SUN Yongcong^1,2, ZENG Xianhua^1,2

1. College of Computer Science and Technology, Chongqing University of Posts and Telecommunication, Chongqing 400065, China;
2. Chongqing Key Laboratory of Computation Intelligence, Chongqing University of Posts and Telecommunications, Chongqing 400065, China

关键词:

图正则化; 字典对学习; 几何近邻关系; 图像分类; 轻度认知功能障碍预测

Keywords:

graph regularization; dictionary pair learning; geometric neighborhood relationship; image classification; mild cognitive impairment prediction

分类号:

TP391;R749

DOI:

10.11992/tis.201709033

摘要:

针对字典对学习（DPL）方法只考虑了同类子字典的重构误差和不同类表示系数的稀疏性，没有考虑图像间的几何近邻拓扑关系的问题。通过近邻保持使得在同类近邻投影系数之间的距离较小，而不同类投影系数之间的距离大，能够有效提高字典对学习算法的分类性能，基于此提出了基于几何近邻拓扑关系的图正则化的字典对学习（GDPL）算法。在ADNI1数据集上对轻度认知功能障碍预测的实验表明，使用GDPL算法学习的编码系数作为特征预测的准确率（ACC）和ROC曲线下的面积（AUC）比使用结合生物标志作为特征预测的准确率提高了2%~6%，使用GDPL算法比DPL算法的实验结果也有提高。

Abstract:

Aiming at dictionary pair learning (DPL) methods only consider the reconstruction error of a sub-dictionary from the same class and the sparseness of coefficients from different classes, and do not consider the geometric neighborhood topological relationships between images. To improve the classification ability of DPL algorithms, we propose a DPL with graph regularization (GDPL) algorithm based on geometric neighborhood topological relationships. This algorithm is based on the idea that keeping the neighborhood relationship makes the distance between the neighborhood projection coefficients of the same kind small, while the distance between projection coefficients of different kinds is large. Experiments on mild cognitive impairment prediction using the ADNI1 dataset show that the coding coefficient learned from the GDPL algorithm is 2%~6% higher than that which uses the combined biomarker as feature prediction, according to accuracy (ACC) and area under curve (AUC) metrics. Moreover, the experimental result obtained using GDPL is also better than that obtained using DPL algorithm.