Staging Liver Fibrosis Based On Texture Features In Gadolinium-Enhanced Magnetic Resonance Imaging

Purpose: We utilized the non-invasivecomputer aided diagnosis(CAD)system to analyze the magnetic resonance(MR)images in a rat model,aiming to improve the accuracy of the liver fibrosis staging dignosis.Methods: A group of 48 healthy male Sprague-Dawley ratswere randomly assigned to theexperimental group(n = 36)and the control group(n = 12).After a week of adaptation,the experimentalrats were subcutaneously injected with a mixture ofcarbon tetrachloride(Baiyin Reagent Factory,Shanghai)and olive oil(50% v/v)at a dose of 0.3 ml/100 g ofbody weight,twice a week for 12 weeks.The first dosewas 0.5 ml/100 g.The control rats were subcutaneously injected with comparable volumes of saline.Each week,between the 4th and 12 th week,rats were randomlyselected for Magnetom Verio Tim 3.0TMR scanning(Siemens)with a rat special coil(3.0T,Chenguangmedicine technology corporation,Shanghai).Then,the equilibriumphase images were acquired post-injection of Magnevist 180 s with 3D VIBE fat-suppressed T1-weightedgradient echo sequence.All rats were grouped based on histological stages(F0-F4).During the process of CAD,the first step was manual section of region of interest(ROI)with 10*10 pixels guided by histology.Next,Lloyd’s algorithm and linear normalization were both used to compress the data to 256 gray level.Then,80 texture features based on the gray level co-occurrence matrix were extracted,80 texture features are autocorrelation,contrast,correlation,cluster prominence,cluster shade,dissimilarity,energy,entropy,homogeneity,maximumprobability,sum average,variance,sum variance,difference variance,sum entropy,difference entropy,information measures of correlation1,informationmeasures of correlation2,inverse difference normalized,and inverse difference moment normalized.Lastly,the back-propagation(BP),Linear,K-nearest neighbor and support vector machine classifiers were individually used to classify F0-F4.Results: Forty-three rats survived after fulfilling the assignedexperimental procedures,five rats died.No fibrosis(F0)was diagnosed in any of the 10 controlled rats,and varying stages of liver fibrosis(F1-4)were scored in 33 experimental rats by histology.There was no F5 stage observed(cirrhosis).Comparable stages of liver fibrosis were detected in all histological specimens acquired from each ofthe 28 experimental rats,but mixed stages of fibrosis amongdifferent specimens were noted in each of the remaining 5 experimental rats,and two adjacent stages of liver fibrosis was found in one rat.With histological staging as a reference,we found that the performance of the BP classifier based on Lloyd’s algorithm in staging fibrosis was better than that of linear normalization.Staging accuracy rates of the former were 67.27%,80.26% and 79.41% for F0 versus F3,F2 versus F4 and F3 versus F4,respectively,secondary to none of the other classifiers.Based on non-linearnormalization,the performance of BP network classifier is better than Linear,K-NN method and SVM.The performance of the linear classifierthat used non-linear normalization was best for F0 vs F4 with 67.16% accuracy,while the K-NNclassifier was best for F2 vs F4 with 69.73%accuracy,and the SVM classifier was best for F3 vs F4 with 73.53% accuracy.Conclusions:1.The performance of the BP classifier with non-linearnormalization in staging fibrosis was better than thatof linear normalization.Based on non-linearnormalization,the performance of BP network classifier is better than Linear,K-NN method and SVM.2.The non-invasive,objective and quantitative CAD system,increased the accuracy of staging liver fibrosis,provided a new idea for the clinical diagnosis of staging of liver fibrosis.