Computer assisted characterization of liver tissue using image texture analysis techniques on B-scan images

In this study, the classification of B-scan ultrasonic liver images using image texture analysis techniques is investigated. The texture analysis algorithms used were the Gray Level Difference Statistics (GLDS), the Gray Level Run Length Statistics (RUNL), the Spatial Gray Level Dependence Matrices (SGLDM) and the Fractal Dimension Texture Analysis (FDTA). All four techniques were applied on four sets of ultrasonic liver images: normal, fatty, cirrhosis and hepatoma. A total of 120 cases were investigated (30 from each class), with all abnormal cases being histologically proven. In each image, a 32/spl times/32 pixel rectangular region-of-interest was selected by an expert physician. Results were classified using the K-nearest neighbor (K-NN) classifier. The FDTA and SGLDM algorithms were able to classify the four sets with an overall accuracy of 78,3% and 77,5% respectively while the RUNL algorithm achieved 74.2% and the GLDS algorithm 70.8% overall accuracy. Combination of RUNL, SGLDM and FDTA improved the overall accuracy to 80%.