A New Look at Gray-level Co-occurrence for Multi-scale Texture Descriptor with Applications to Characterize Colorectal Polyps via Computed Tomographic Colonography

Characterizing colon polyps is clinically important but technically challenging. The gray-level co-occurrence matrix (GLCM)-based texture descriptor, proposed by Haralick et al., has shown the potential to relive the challenging. This study aims to increase the potential by exploring multiple-displacement GLCM descriptor (MDGLCM), multiple-stride GLCM descriptor (MSGLCM) and adaptive-sampling GLCM descriptor (ASGLCM). Both MDGLCM and MSGLCM use multiple step shifts to increase the texture information based on the Haralick model. ASGLCM investigates adaptive sampling on both direction and displacement for the purpose of increasing the texture patterns and minimizing the spatial variation and is the main contribution of this work. This method integrates the ranked texture descriptors via eliminating the redundant information to characterize 63 polyp masses, including 32 invasive adenocarcinoma and 31 benign adenomas. For comparison purpose, the texture descriptor from the Haralick model was implemented (in the same manner as the above presented texture descriptors) as baseline, which predicted the lesions by AUC (area under the curve of receiver operating characteristics) score of 0.8326 and standard deviation 0.0646. The ASGLCM improved the prediction power to 0.9023 with standard deviation 0.0362, and the improvement is statistically significant.