Liu Meizhu, Lu Le, Ye Xiaojing, Yu Shipeng, Salganicoff Marcos
University of Florida, Gainesville, FL 32611, USA.
Med Image Comput Comput Assist Interv. 2011;14(Pt 3):41-8. doi: 10.1007/978-3-642-23626-6_6.
Classification is one of the core problems in computer-aided cancer diagnosis (CAD) via medical image interpretation. High detection sensitivity with reasonably low false positive (FP) rate is essential for any CAD system to be accepted as a valuable or even indispensable tool in radiologists' workflow. In this paper, we propose a novel classification framework based on sparse representation. It first builds an overcomplete dictionary of atoms for each class via K-SVD learning, then classification is formulated as sparse coding which can be solved efficiently. This representation naturally generalizes for both binary and multiwise classification problems, and can be used as a standalone classifier or integrated with an existing decision system. Our method is extensively validated in CAD systems for both colorectal polyp and lung nodule detection, using hospital scale, multi-site clinical datasets. The results show that we achieve superior classification performance than existing state-of-the-arts, using support vector machine (SVM) and its variants, boosting, logistic regression, relevance vector machine (RVM), or kappa-nearest neighbor (KNN).
分类是通过医学图像解读进行计算机辅助癌症诊断（CAD）的核心问题之一。对于任何CAD系统而言，要想在放射科医生的工作流程中被视为有价值甚至不可或缺的工具，具备高检测灵敏度且假阳性（FP）率合理至关重要。在本文中，我们提出了一种基于稀疏表示的新型分类框架。它首先通过K-SVD学习为每个类别构建一个过完备原子字典，然后将分类问题转化为可高效求解的稀疏编码问题。这种表示方法自然地适用于二分类和多分类问题，既可以用作独立的分类器，也可以与现有的决策系统集成。我们的方法在用于结肠息肉和肺结节检测的CAD系统中，使用医院规模的多站点临床数据集进行了广泛验证。结果表明，与使用支持向量机（SVM）及其变体、提升算法、逻辑回归、相关向量机（RVM）或κ-最近邻（KNN）的现有最先进方法相比，我们取得了更优异的分类性能。
