Improving Colon Histopathology Classification Using Graph-Based Structural Features and HOG Descriptors

Abstract

Reliable classification of colon histopathology images plays an important role in the early detection and diagnosis of colorectal cancer. However, conventional microscopic examination is time-consuming and may be affected by inter-observer variability. This study proposes an interpretable hybrid computational framework that combines Histogram of Oriented Gradients (HOG) features with graph-theoretic descriptors derived from the spatial organization of cell nuclei. To capture complementary structural information, four graph-based representations: Delaunay triangulation, Minimum Spanning Tree (MST), k-Nearest Neighbor (k-NN), and Voronoi diagrams, are systematically examined. The extracted structural and texture features are integrated and used to train a Random Forest classifier. Experiments were performed on an augmented dataset containing approximately 5,000 images per class. The results indicate that the proposed feature-fusion approach provides improved classification accuracy, stability, and sensitivity for cancer detection compared with the use of individual feature descriptors. By combining image texture with interpretable spatial and structural information, the proposed framework offers a balanced approach in terms of predictive performance, interpretability, and computational efficiency, highlighting its potential for computer-assisted analysis of colon histopathology images.