Quantitative Diagnosis Of Liver Fibrosis Stage And Collagen Proportionate Area Based On Deep Learning Model Of Enhanced CT

Part 1 Prediction of liver fibrosis staging based on deep learning model of enhanced CTObjective:Developed and validtaed the deep learning model based on enhanced CT in staging liver fibrosisMaterials and methods:452 cases in Zhujiang Hospital,Southern Medical University and Renji Hospital,Shanghai Jiaotong University School of Medicine were collected respectively from January 2017 to December 2020.366 cases in one center were used for model construction and validation,and 86 cases in another center were used for external validation.Two pre-trained nnUnet segmentation models were used to segment the liver and intrahepatic tumors respectively.For cases with liver space-occupying lesions,two Area of interest were subtracted to obtain the final whole liver ROI,For cases without liver space-occupying lesions,the segmented liver ROI was directly used as the final ROI.The 2D Efficient Net-B0 end-to-end classification network was used to construct the deep learning model of liver fibrosis staging based on no-contrast scan CT,arterial phase,portal phase,equilibrium phase and combined four-phased CT.Spearman correlation analysis was used to test the correlation between fibrosis stage and deep learning model score(DL-Score).Single factor analysis was used to evaluate the difference of DL-Score in different stages of liver fibrosis.ROC analysis was used to test the diagnostic performance of the deep learning model,and Delong test was used to compare the differences of the diagnostic performance of the models,P<0.05 was considered statistically significant.Results:DL-Score of deep learning model score was significantly positively correlated with liver fibrosis stage.In univariate analysis,the DL-Score of each model had statistically significant differences in each phase of liver fibrosis.In the nocontrast scan model,the AUC of the training set/internal validation set/external validation set were 0.998/0.947/0.743(FI-4),0.998/0.968/0.742(F2-4),0.998/0.938/0.678(F3-4)and 0.998/0.900/0.752(F4),respectively.In the arterial phase model,the AUC of the training set/internal validation set/external validation set were 0.994/0.953/0.762(F1-4),0.998/0.977/0.676(F2-4),0.997/0.968/0.616(F3-4)and 0.996/0.924/0.673(F4),respectively.In the portal venous phase model,the AUC of the training set/internal validation set/external validation set were 0.973/0.926/0.632(F1-4),0.970/0.982/0.632(F2-4),0.952/0.918/0.661(F3-4)and 0.940/0.885/0.678(F4),respectively.In the equilibrium phase model,the AUC of the training set/internal validation set/external validation set were 0.971/0.906/0.715(F1-4),0.968/0.947/0.711(F2-4),0.936/0.868/0.647(F3-4)and 0.926/0.842/0.666(F4),respectively.In the combined 4-phased model,the AUC of the training set/internal validation set/external validation set were 0.997/0.944/0.712(F1-4),0.998/0.959/0.687(F2-4),0.999/0.891/0.704(F3-4)and 0.999/0.879/0.760(F4),respectively.In Delong test,there was no significant difference in the diagnostic performance of combined 4-phased model compared with each single-phase deep learning model and each single-phase model.Conclusion:The deep learning model based on enhanced CT can be used for liver fibrosis staging,and the diagnostic efficiency of the combined 4-phased model is not better than that of the single-phase deep learning diagnostic model.Part 2 Deep learning model based on enhanced CT predicts collagen proportionate areaObjective:To preliminarily explored the feasibility of quantitative diagnosis of CPA stage based on the deep learning model of enhanced CT.Materials and methods:The research object and inclusion criteria were the same as above.Collagen proportional area was measured by digital image analysis of the patient’s pathological samples.The data set for model construction and validation was the same as above.The 2D EfficientNet-B0 end-to-end classification network was used to construct the deep learning model.The model evaluation and validation methods were the same as above.Results:Liver fibrosis stage was associated with CPA(R=0.650,P<0.001).In training set,the deep learning model scores based on CPA Stage(DLcPA-Score)in each period was significantly positively correlated with CPA stage.In the internal validation set,except for the no-contrast scan and combined 4-phased model,the DLcPA-Score of the remaining models was significantly positively correlated with the CPA stage;in the external validation set,except for the arterial phase model,theDLcPA-Score of the other models was significantly positively correlated with the CPA stage.In univariate analysis,the DLcPA-Score of each model in each CPA stage were statistically significant.In the no-contrast scan model,the AUC of the training set/internal validation set/external validation set were 0.999/0.620/0.662(C2-4),1.000/0.641/0.569(C3-4)and 1.000/0.611/0.835(C4),respectively.In the arterial phase model,the AUC of the training set/internal validation set/external validation set were 0.905/0.697/0.595(C2-4),0.901/0.693/0.520(C3-4)and 0.924/0.722/0.662(C4),respectively.In the portal venous phase model,the AUC of the training set/internal validation set/external validation set were 0.848/0.726/0.683(C2-4),0.848/0.686/0.634(C3-4)and 0.852/0.750/0.622(C4),respectively.In the equilibrium phase model,the AUC of the training set/internal validation set/external validation set were 0.820/0.731/0.698(C2-4),0.822/0.712/0.656(C3-4)and 0.832/0.745/0.627(C4),respectively.In the combined 4-phased model,the AUC of the training set/internal validation set/external validation set was 0.804/0.688/0.750(C2-4),0.820/0.686/0.664(C3-4)and 0.814/0.690/0.700(C4),respectively.In the pairwise comparison of each model,the diagnostic performance of the no-contrast scan model in the training set was better than that of other models,and there was no significant difference in the diagnostic performance of each model in the internal and external validation sets.Conclusion:Liver fibrosis stage was correlated with CPA.The deep learning model has moderate classification prediction ability for CPA staging,and the diagnostic efficiency of the combined 4-phased model is not better than that of the single-phase deep learning diagnostic model.