Texture analysis of multi-phase MRI images to detect expression of Ki67 in hepatocellular carcinoma

Abstract

To determine whether texture analysis of preoperative magnetic resonance imaging (MRI) images could be used to detect Ki67 expression, a widely used cell proliferation marker in hepatocellular carcinoma (HCC). In total, 83 patients were included, 25 with low Ki67 (Ki67 ≤10%) HCC expression and 58 with high Ki67 (Ki67 ≥10%) HCC expression as demonstrated by retrospective surgical evaluation. All patients were examined using a 3 T MRI unit with one standard protocol. The region of interest was drawn manually by one radiologist. Texture analysis included histogram, co-occurrence matrix, run-length matrix, gradient, auto-regressive model, and wavelet transform features as calculated by MaZda (version 4.6; quantitative texture analysis software). The features reduced by the Fisher, probability of classification error, and average correlation coefficient (POE+ACC), mutual information were used to select the features that predicted Ki67 proliferation status with highest accuracy and then using the B11 program for data analysis and classification. The misclassification rate of the principal component analysis (PCA) in the hepatobiliary phase (HBP), T2-weighted imaging (T2WI), arterial phase (AP), and portal vein phase (PVP) was 36/83 (43.37%), 35/82 (42.68%), 40/83 (48.19%), and 34/83 (40.96%), respectively. The misclassification of the linear discriminant analysis in HBP, T2WI, AP, and PVP phase was 13/83 (15.66%), 21/82 (25.61%), 9/83 (10.84%), and 8/83 (9.64%), respectively. The misclassification of the nonlinear discriminant analysis in HBP, T2WI, AP, and PVP phase was 7/83 (8.43%), 6/82 (7.32%), 5/83 (6.02%), and 7/83 (8.43%), respectively. Texture analysis of HBP, AP, and PVP were helpful for predicting Ki67 expression and may provide a less-invasive method to investigate critical histopathology markers for HCC.