Discriminative feature selection by optimal manifold search for neoplastic image recognition

Abstract

An endocytoscope provides ultramagnified observation that enables physicians to achieve minimally invasive and real-time diagnosis in colonoscopy. However, great pathological knowledge and clinical experiences are required for this diagnosis. The computer-aided diagnosis (CAD) system is required that decreases the chances of overlooking neoplastic polyps in endocytoscopy. Towards the construction of a CAD system, we have developed texture-feature-based classification between neoplastic and non-neoplastic images of polyps. We propose a feature-selection method that selects discriminative features from texture features for such two-category classification by searching for an optimal manifold. With an optimal manifold, where selected features are distributed, the distance between two linear subspaces is maximised. We experimentally evaluated the proposed method by comparing the classification accuracy before and after the feature selection for texture features and deep-learning features. Furthermore, we clarified the characteristics of an optimal manifold by exploring the relation between the classification accuracy and the output probability of a support vector machine (SVM). The classification with our feature-selection method achieved 84.7% accuracy, which is 7.2% higher than the direct application of Haralick features and SVM.