Effective staging of fibrosis by the selected texture features of liver: Which one is better, CT or MR imaging?

Abstract

Purpose: Texture patterns of hepatic fibrosis are one of the important biomarkers to diagnose and classify chronic liver disease from initial to end stage on computed tomography (CT) or magnetic resonance (MR) images. Computer-aided diagnosis (CAD) of liver cirrhosis using texture features has become popular in recent research advances. To date, however, properly selecting effective texture features and image parameters is still mostly undetermined and not well-defined. In this study, different types of datasets acquired from CT and MR images are investigated to select the optimal parameters and features for the proper classification of fibrosis.

Methods: A total of 149 patients were scanned by multi-detector computed tomography (MDCT) and 218 patients were scanned using 1.5T and 3T superconducting MR scanners for an abdominal examination. All cases were verified by needle biopsies as the gold standard of our experiment, ranging from 0 (no fibrosis) to 5 (cirrhosis). For each case, at least four sequenced phase images are acquired by CT or MR scanners: pre-contrast, arterial, portal venous and equilibrium phase. For both imaging modalities, 15 texture features calculated from gray level co-occurrence matrix (GLCM) are extracted within an ROI in liver as one set of input vectors. Each combination of these input subsets is checked by using support vector machine (SVM) with leave-one-case-out method to differentiate fibrosis into two groups: non-cirrhosis or cirrhosis. In addition, 10 ROIs in the liver are manually selected in a disperse manner by experienced radiologist from each sequenced image and each of the 15 features are averaged across the 10 ROIs for each case to reduce the validation time. The number of input items is selected from the various combinations of 15 features, from which the accuracy rate (AR) is calculated by counting the percentage of correct answers on each combination of features aggregated to determine a liver stage score and then compared to the gold standard.

Results: According to the accuracy rate (AR) calculated from each combination, the optimal number of texture features to classify liver fibrosis degree ranges from 4 to 7, no matter which modality was utilized. The overall performance calculated by the average sum of maximum AR value of all 15 features is 66.83% in CT images, while 68.14%, and 71.98% in MR images, respectively; among the 15 texture features, mean gray value and entropy are the most commonly used features in all 3 imaging datasets. The correlation feature has the lowest AR value and was removed as an effective feature in all datasets. AR value tends to increase with the injection of contrast agency, and both CT and MR images reach the highest AR performance during the equilibrium phase.

Conclusions: Comparing the accuracy of classification with two imaging modalities, the MR images have an advantage over CT images with regards to AR performance of the 15 selected texture features, while 3T MRI is better than 1.5T MRI to classify liver fibrosis. Finally, the texture analysis is more effective during equilibrium phase than in any of the other phased images.