Deep Learning Model Of Diagnosis And Staging Liver Fibrosis In Chronic Hepatitis B By Using MR Images

ObjectiveTo research the performance of deep learning model of automatically segments region of interest,and diagnosis and staging liver fibrosis in chronic hepatitis B by using multiple magnetic resonance sequences.MethodsCollected from January 2015 to December 2018,there was accumulated total of 304 chronic hepatitis B patients with clinical data(gender,age,etc.),serological examination(blood routine,HBV DNA,detection of HBV markers,liver function,etc.),liver biopsy pathology results and standardization scanning of magnetic resonance imaging(T2 weighted images and fat-suppression T2 weighted images,and T1 weighted images enhancement scanning delayed phase)completed within 7 days before and after liver biopsy.The data aged 40.34±11.60 years,ranging from 16 to 86 years,including 213 males and 91 females.Two radiologists manually annotated region of interests of hepatic segmentⅥ.Input the magnetic resonance images to the deep learning model,automatic annotated region of interests of hepatic segment VI,comparing the results of automatic segmentation and manual segmentation.We randomly sampled the patients into train group(213 cases)and the test group(91 cases)with a ratio of 7:3.The train group were used to training the deep learning convolutional neural network model and the test group were used to test the diagnostic efficiency of the model for liver fibrosis stage.The diagnostic efficacy of only index model,only image model,image-index combined model and single sequence model and multiple sequences model in diagnosis of liver fibrosis in chronic hepatitis B was evaluated by receiver operating characteristic.Area under the curve,accuracy,sensitivity and specificity in the diagnosis of advanced fibrosis(>=S2)and significance fibrosis(>=S3)were calculated and compared.ResultsThe automatic-segmentation of hepatic segment VI by our pipeline,app roximately 2000 slices can be automatically split segmented in 45 seconds,and our pipeline managed to segment hepatic segment VI by multiple seque nces with dice score of 0.85.In advanced fibrosis(>=S2),the AUC value s of only index,only MR image and image-index combined model were 0.7466,0.6654 and 0.7810,respectively,with accuracy of 0.7912,0.6143 and 0.7285,sensitivity of 0.5152,0.6146 and 0.7289,and specificity of 0.9483,0.6139 and 0.7281,respectively.In significance fibrosis(>=S3),t he AUC values of the individual clinical indicators,the individual MR im ages,and the combined clinical indicators-MR images were 0.6618,0.5516,and 0.7544,respectively;the accuracy was 0.6703,0.5312,and 0.6909,re spectively;the sensitivity was 0.6935,0.5321,and 0.6912,respectively;and the specificity was 0.6207,0.5309,and 0.6908,respectively.The AU C of the clinical indicators-MR image model was the highest in the three models.In advanced fibrosis stage(>=S2),single sequence model(T2 weig hted and T2 weighted pressure grease series,and T1 weighted enhancement scanning delay)and multiple sequences model of AUC values were 0.7902,0.7576,0.6999,0.7810,and the accuracy were 0.7412,0.7061,0.6472,0.7285,and the sensitivity were 0.7418,0.7087,0.6475,0.7289,and the spec ificity were 0.7389,0.6974,0.6462,0.7281;In significance fibrosis(>=S3),single sequence model(T2 weighted and T2 weighted pressure grease series,and T1 weighted enhancement scanning delay)and multiple sequenc es fusion model of AUC values were 0.7330,0.7010,0.7578,0.7544,and t he accuracy were 0.6641,0.6462,0.6804,0.6909,and the sensitivity were 0.6654,0.6479,0.6892,0.6912,and the specificity of 0.6635,0.6454,0.6880,0.6908,respectively.The AUC,accuracy,sensitivity and specificit y of T2-weighted lipid compression sequences were all lower than those of T2-weighted and T1-weighted enhanced scan and multiple sequences fusion models.ConclusionOur pipeline can improve the efficiency of ROI segmentation.The deep learning convolutional neural network model based on multiple magnetic resonance sequences has certain diagnostic value for the classification of liver fibrosis in chronic hepatitis B,and has certain distinguishing ability for early liver fibrosis,which can be used as an auxiliary means for the selection of treatment plan and the evaluation of curative effect in patients with chronic hepatitis B.