Three-dimensional CT texture analysis of anatomic liver segments can differentiate between low-grade and high-grade fibrosis

Abstract

Abstract Background CT texture analysis (CTTA) has been successfully used to assess tissue heterogeneity in multiple diseases. The purpose of this work is to demonstrate the value of three-dimensional CTTA in the evaluation of diffuse liver disease. We aimed to develop CTTA based prediction models, which can be used for staging of fibrosis in different anatomic liver segments irrespective of variations in scanning parameters. Methods We retrospectively collected CT scans of thirty-two chronic hepatitis patients with liver fibrosis. The CT examinations were performed on either a 16- or a 64-slice scanner. Altogether 354 anatomic liver segments were manually highlighted on portal venous phase images, and 1117 three-dimensional texture parameters were calculated from each segment. The segments were divided between groups of low-grade and high-grade fibrosis using shear-wave elastography. The highly-correlated features (Pearson r > 0.95) were filtered out, and the remaining 453 features were normalized and used in a classification with k-means and hierarchical cluster analysis. The segments were split between the train and test sets in equal proportion (analysis I) or based on the scanner type (analysis II) into 64-slice train 16-slice validation cohorts for machine learning classification, and a subset of highly prognostic features was selected with recursive feature elimination. Results A classification with k-means and hierarchical cluster analysis divided segments into four main clusters. The average CT density was significantly higher in cluster-4 (110 HU ± SD = 10.1HU) compared to the other clusters (c1: 96.1 HU ± SD = 11.3HU; p < 0.0001; c2: 90.8 HU ± SD = 16.8HU; p < 0.0001; c3: 93.1 HU ± SD = 17.5HU; p < 0.0001); but there was no difference in liver stiffness or scanner type among the clusters. The optimized random forest classifier was able to distinguish between low-grade and high-grade fibrosis with excellent cross-validated accuracy in both the first and second analysis (AUC = 0.90, CI = 0.85–0.95 vs. AUC = 0.88, CI = 0.84–0.91). The final support vector machine model achieved an excellent prediction rate in the second analysis (AUC = 0.91, CI = 0.88–0.94) and an acceptable prediction rate in the first analysis (AUC = 0.76, CI = 0.67–0.84). Conclusions In conclusion, CTTA-based models can be successfully applied to differentiate high-grade from low-grade fibrosis irrespective of the imaging platform. Thus, CTTA may be useful in the non-invasive prognostication of patients with chronic liver disease.

Keywords

• Machine learning
• Texture analysis
• Liver fibrosis
• Computed tomography
• Prediction model
• Radiomics