Staging Liver Fibrosis Based On Computer-aided Diagnosis System In Routine Magnetic Resonance Imaging

ObjectiveThis study aimed to create a computer-aided diagnosis system to analyze MR routine scan images in a rabbit model,in order to explore the value of machine learning model combining semi-supervised learning and active learning in staging diagnosis of liver fibrosis.Methods1.Establishment of animal model of liver fibrosis,image acquisition and pathologyThirty-five healthy male rabbits were divided into experimental group（n=30）and control group（n=5）to establish liver fibrosis model.Rabbits in the experimental group had received subcutaneous injections at neck twice a week.We used the drug which comprising 50%carbon tetrachloride（CCl4）in olive oil（0.1m L/kg at 1-3 weeks,0.2m L/kg at 4-6 weeks,0.3m L/kg at 7-10 weeks）.Rabbits in control group were injected with the same dose of normal saline.At the end of 5,6,7 and 10 weeks after injection,the rabbits were scanned by 3.0T MR and the sequence images of T₁WI-FS and T₂WI-FS were obtained.The rabbits were sacrificed after MR scanning,and their livers were stained by HE and Masson,respectively.The rabbits were grouped according to the Metavir scoring system:normal liver tissue（F0）,mild fibrosis（F1）,moderate fibrosis（F2）,severe fibrosis（F3）,and early cirrhosis（F4）.Then they were divided into three groups:early liver fibrosis group（F1-2）,advanced liver fibrosis group（F3-4）,and normal liver group（F0）.2.Classification and recognition system based on computer aided diagnosis（1）Region of interest extractionOn the sequence images of T₁WI-FS and T₂WI-FS,we drew a total of 360 square ROIs of 20×20 pixels.180 ROIs were obtained for each sequence images,and their positions correspond to each other.including normal liver tissue（F0,n=32）,mild fibrosis（F1,n=37）,moderate fibrosis（F2,n=33）,severe fibrosis（F3,n=54）,early cirrhosis（F4,n=24）,early liver fibrosis（F1-2,n=70）,and advanced liver fibrosis（F3-4,n=78）.（2）Texture feature extraction and selectionAfter image preprocessing,the gray-level co-occurrence matrix（GLCM）was used to calculate the texture characteristic values of ROI image in four directions（0°,45°,90°,135°）.Then the average value was obtained to select the parameter features,and finally a 16-dimensional feature was obtained for each ROI.Specific parameters were as follows:power,contrast ratio,inverse difference moment,variance,entropy,inertia moment,correlation,sum average,sum entropy,sum variance,difference entropy,difference variance,information measures of correlation 1and 2,mean gray value,standard deviation.（3）Machine Learning model establishmentA label propagation algorithm combining semi-supervised learning and active learning was used to build a classifier model.Five groups of ROIs were classified and identified respectively,and then three groups of ROIs were compared between groups.ResultsA total of 33 rabbits completed the entire experimental process.The pathological results showed that the stages of liver fibrosis in each liver lobe of all surviving rabbits were not completely consistent,so the rabbits as a whole were ignored and each liver lobe was taken as the research object to plot ROI.Finally,the classification and identification results were as follows:1.Classification and recognition results based on T₁WI-FS imagesThe average accuracy of five-stage classification and recognition（5 groups of ROIs）was 60.6%,and the accuracy of F0,F2 and F3 liver fibrosis was 87.5%,66.7%and 64.8%,respectively.The average accuracy of pairwise classification and recognition（3 groups of ROIs）reached more than 90%in the identification of F0 and F1-4,F0 and F1-2,F0 and F3-4,among which F0 and F3-4 were the best,with the accuracy of 96.4%.2.Classification and recognition results based on T₂WI-FS imagesThe recognition performance of the classifier is not stable.The average accuracy of five-stage classification and recognition（5 groups of ROIs）was 44.4%,The average accuracy of pairwise classification and recognition（3 groups of ROIs）was ineffective because of the high misjudgment rate of F0 phase.The average accuracy of identifying F1-2 and F3-4 was 69.6%,and the accuracy of distinguishing F3-4 was 83.3%.Conclusion1.The emerging SSL and AL classifier model constructed based on MR routine scan image could be applied to the classification and recognition of rabbit liver fibrosis model.The classification and recognition effect of T₁WI-FS images was better than that of T₂WI-FS images.Pairwise classification and identification was significantly better than five-stage classification and identification,could effectively distinguish the normal liver tissue,early liver fibrosis,advanced liver fibrosis.2.Non-invasive computer-aided diagnosis technology has certain value in staging diagnosis of liver fibrosis,which may provide the potential application value for clinical diagnosis and treatment and prognosis in the future.