Prediction Of Microvascular Invasion In Hepatocellular Carcinoma Using Convolutional Neural Network Based On IVIM-DWI And Gd-EOB-DTPA Enhanced MRI

Part 1 Prediction of microvascular invasion in hepatocellular carcinoma using convolutional neural network based on IVIM-DWIObjectives:This study aimed to explore the diagnostic performance of intravoxel incoherent motion(IVIM)diffusion weighted imaging for prediction of microvascular invasion(MVI)in hepatocellular carcinoma(HCC)using convolutional neural network(CNN).Methods:This study included 114 patients with pathologically confirmed HCC.All patients underwent MRI examination including IVIM sequence with 9b-values preoperatively.HCC patients were divided into MVI positive and MVI negative groups according to postoperative pathological results.114 HCC patients were randomly divided into a training cohort(n=74;28 MVI-positive and 46 MVI-negative)and a validation cohort(n=40;15 MVI-positive and 25 MVI-negative).Clinicopathological information of all patiente,including age,sex,hepatitis B status(positive or negative),α-fetoprotein(AFP)level,aspartate transaminase(AST),alanine transaminase(ALT),bilirubin,albumin,platelet count,tumor size and Edmondson-Steiner grade,were retrieved from the electronic medical record system and pathological reports,and then the characteristics related to MVI in HCC patients were analyzed.Then,we analyzed the IVIM parameter values associated with the MVI in HCC patients.All IVIM-DWI images were transferred to a workstation for postprocessing to automatically generate ADC,D,D*and f parameter maps.The regions of interest(ROIs)were drawn manually by a radiologist on DWI images with b1000 using publicly available software ImageJ,the ROI of each patient was placed on a maximum representative slice in HCC.Then,the ADC,D,D*and f values were calculated using parameter maps with the outlined ROIs by ImageJ.The emphasis of this study was to develop deep learning model based on IVIM-DWI.First,we extracted tumor areas from original images using MATLAB and all tumor areas were normalized to 32 × 32.Then,tumor areas from 9 b-value images were superimposed in the channel dimension to obtain a b-value volume with a shape of(32,32,9)dimension to use as the input of the CNN.And then,data augmentation was performed by randomly rotating the samples of the training cohort around the center point 50 times.Finally,deep features to predict MVI in HCC were directly derived from a b-value volume based on the CNN.Moreover,a deep learning model based on parameter maps and a fusion model combined with deep features of IVIM,clinical characteristics and IVIM parameters were also constructed.The area under curve(AUC),accuracy,sensitivity and specificity of each deep learning model in the validation cohort were evaluated.Results:Among clinical characteristics,the AFP level was also significantly different between MVI positive and MVI negative groups in the training(P=0.031)and validation cohorts(P=0.011).Tumor size differed significantly in the training cohort(P=0.016)but was not confirmed in the validation cohort(P=0.074).Among the IVIM parameter values,only the ADC value had a statistically significant difference between the MVI positive and MVI negative groups in the validation cohort(P=0.005).For the deep learning models,deep features directly extracted from IVIM-DWI[0.810(range 0.760,0.829)]using CNN for prediction of MVI yielded better performance than those from IVIM parameter maps[0.590(range 0.555,0.643)].Furthermore,the performance of the fusion model combined with deep features of IVIM-DWI,clinical features(AFP level and tumor size)and ADC value was slightly improved.The AUC,accuracy,sensitivity,and specificity of the fusion model were 0.829(range,0.776,0.848),0.775(range,0.700,0.800),0.666(range,0.600,0.860),and 0.870(range,0.760,0.920),respectively.Conclusions:Deep learning with CNN based on IVIM-DWI can be conducive to preoperative prediction of MVI in patients with HCC.Part 2 Prediction of microvascular invasion in hepatocellular carcinoma using convolutional neural network based on Gd-EOB-DTPA enhanced MRIObjectives:The aim of this study was to investigate the diagnostic performance of convolutional neural network(CNN)based on Gd-EOB-DTPA enhanced magnetic resonance imaging(MRI)for prediction of microvascular invasion(MVI)in hepatocellular carcinoma(HCC).Methods:This retrospective study enrolled 137 patients with pathologically confirmed HCC.All patients underwent Gd-EOB-DTPA enhanced MRI preoperatively.According to postoperative pathological results,HCC patients were divided into MVI positive group and MVI negative group.137 HCC patients were randomly divided into a training cohort(n=96;35 MVI-positive and 61 MVI-negative)and a validation cohort(n=41;14 MVI-positive and 27 MVI-negative).Clinical information of all patients,including age,sex,hepatitis B status(positive or negative),α-fetoprotein(AFP)level,aspartate transaminase(AST),alanine transaminase(ALT),bilirubin,albumin,platelet count and degree of tumor differentiation,were retrieved from the electronic medical record system and pathological reports.All MR images were analyzed by two radiologists independently.The imaging features include tumor size,tumor margin,enhancement pattern,tumor capsule,arterial peritumoral enhancement,tumor hypointensity on hepatobiliary phase(HBP)images and peritumoral hypointensity on HBP.Clinical information and imaging features were compared between the MVI positive group and MVI negative group in the training cohort and validation cohort.For statistically significant features in both training cohort and validation cohort,its efficiency in predicting MVI was evaluated by receiver operating characteristic curve(ROC curve).The emphasis of this study was to develop deep learning models based on Gd-EOB-DTPA enhanced MRI.First,we extracted tumor areas from original images using MATLAB and all tumor areas were normalized to 32 × 32 × 32.Then,the random flip method was used for data enhancement.Finally,the ResNet model was used to extract the deep features from the Gd-EOB-DTPA enhanced MRI for MVI prediction.The area under curve(AUC),accuracy,sensitivity and specificity of each deep learning model in the validation cohort were evaluated.Results:For clinical information and imaging features,only tumor size and tumor margin had statistically significant differences between the MVI-positive group and MVI-negative group in both the training cohort and validation cohort.The AUCs(95%CI)of tumor size and tumor margin were 0.706(95%CI 0.543,0.870)and 0.728(95%CI 0.559,0.896),respectively.For the deep learning model based on the Gd-EOB-DTPA enhanced MRI of each phase,the deep learning model based on the images of the transitional phase and the hepatobiliary phase showed better performance for MVI prediction,with AUCs of 0.786(0.775,0.798)and 0.838(0.768,0.860),respectively.In addition,the MVI prediction performance of the fusion model based on the transitional and hepatobiliary phase images was better than that of the deep learning model based on single-phase Gd-EOB-DTPA enhanced MRI,and its AUC,accuracy,sensitivity and specificity were 0.878(0.867,0.993),0.889(0.866,0.993),0.812(0.765,0.878)and 0.945(0.894,0.950),respectively.Conclusions:Deep learning with CNN based on Gd-EOB-DTPA enhanced MRI can be conducive to preoperative prediction of MVI in patients with HCC.