Significant Feature Dimensionality Reduction for Histopathology Image Retrieval as a Tool for Healthcare Decisions Support

Content-Based Histopathology Image Retrieval (CBHIR) is often used as a powerful tool to support decision making by clinicians, and facilitate the disease detection and diagnosis. In CBHIR, feature extraction plays a key role for representing and interpreting histopathological images of several tissues. To enhance the retrieval performance, combining several features is commonly used in CBHIR; however, the high-dimension of the combined features often makes the process of retrieving very time-consuming. To resolve this challenge, we propose in this paper a method for reducing the dimension of a combined feature vector formed of several features extracted using the Improved local binary patterns (ILBP) descriptor, and coded in different color spaces. To reduce the size of the combined color ILBP feature, we adopt a selection scheme for picking only the most important patterns of the latter according to optimal weights associated to the patterns. The optimal weights are generated using the Particle Swarm Optimization (PSO) algorithm. The results of the experiments performed on Kimia Path960 dataset confirm the efficiency of the proposed method in reducing the feature dimensionality and the retrieval time while maintaining as good as or better retrieval performance than the combined feature.