A Texture Analysis Method Based on Statistical Contourlet Coefficient Applied to the Classification of Pancreatic Cancer and Normal Pancreas

Abstract

Purpose: To explore the value of a texture analysis method based on statistical contourlet coefficient in the computer-aided diagnosis of pancreatic cancer and normal pancreas. Methods: This paper proposed a texture analysis method based on statistical contourlet coefficient (SCC) to extract the quantitative features of regions of interest (ROIs) in non-enhanced CT images. The SCC method consisted of two steps. First, it decompose an ROI into several subbands at multiple directions in multiple layers, where a "9-7" filter was applied in the Laplacian pyramid filtering stage and a "pkva" filter was applied in the directional filtering stage. Then, it performed normalization on the coefficient matrices of the subbands and extracted the first and second order statistical features of the normalized matrices. Six traditional texture analysis methods that are widely used for medical image processing were used for comparisons. After the feature extraction, feature selection and classification (10-fold cross training and test) were performed, and the classification results were evaluated. Results: The proposed method achieved the best classification result: the average accuracy was 79.52%; the average sensitivity was 78.5%; the average specificity was 80.63%; the average AUC was 0.848. Conclusions: It indicates that the texture analysis method based on statistical contourlet coefficient is rewarding for computer-aided diagnosis of pancreatic cancer and normal pancreas using non-enhanced CT images. It can reduce the workload of radiologists and play a significant guiding effect on junior radiologists.