High-order graph matching kernel for early carcinoma EUS image classification

Abstract

Endoscopic ultrasonography (EUS) is limited by variability in the examiner's subjective interpretation to differentiate between normal, leiomyoma of esophagus and early esophageal carcinoma. By using information otherwise discarded by conventional EUS systems, quantitative spectral analysis of the raw pixels (picture elements) underlying EUS image enables lesions to be characterized more objectively. In this paper, we propose to represent texture features of early esophageal carcinoma in EUS images as a graph by expressing pixels as nodes and similarity between the gray-level or local features of the EUS image as edges. Then, similarity measurements such as a high-order graph matching kernel can be constructed so as to provide an objective quantification of the properties of the texture features of early esophageal carcinoma in EUS images. This is in terms of the topology and connectivity of the analyzed graphs. Because such properties are directly related to the structure of early esophageal carcinoma lesions in EUS images, they can be used as features for characterizing and classifying early esophageal carcinoma. Finally, we use a refined SVM model based on the new high-order graph matching kernel, resulting an optimal prediction of the types of esophageal lesions. A 10-fold cross validation strategy is employed to evaluate the classification performance. After multiple computer runs of the new kernel SVM model, the overall accuracy for the diagnosis between normal, leiomyoma of esophagus and early esophageal carcinoma was 93 %. Moreover, for the diagnosis of early esophageal carcinoma, the average accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were 89.4 %, 94 %, 95 %, 89 %, and 97 % respectively. The area under all the three ROC curves were close to 1.