| BackgroundHepatocellular carcinoma(HCC)is the fourth leading cause of cancer-related death worldwide.According to the International Agency for Research on Cancer,about83,0180 people died of HCC in 2020,and the number of deaths is expected to exceed 1million in 2025.Chronic hepatitis or cirrhosis is a high-risk factor for HCC,and70%～90% of HCC develops in the background of cirrhosis or progressive hepatic fibrosis.Current international guidelines recommend imaging examinations such as Computed Tomography(CT)and Magnetic Resonance Imaging(MRI)for screening,diagnosis,staging and treatment for HCC in high-risk populations.However,HCC is characterized by insidious onset and significant heterogeneity so that it is difficult for the early detection,diagnosis or making personalized and precise treatment for HCC by naked eyes.Deep learning methods,especially convolutional neural network(CNN),have demonstrated superior performance in medical image analysis tasks by virtue of their powerful feature extraction capabilities.In this study,we explored the value of convolutional neural networks for automatic detection of focal liver lesions(FLLs),HCC diagnosis,and microvascular invasion(MVI)prediction.MethodsExperiment 1 This part explored the effectiveness of automatic detection of FLLs in contrast-enhanced CT images using CNN architectures.2003 patients with 8,597 FLLs were enrolled in this study.We trained detection models based on 2.5D and 3D CNN frameworks using 567 patients with 3892 FLLs and validated on a relatively large independent cohort of 1436 patients with 4723 FLLs.The detection performance across different phases(arterial,portal venous and combined phases)were assessed for the2.5D model.The lesions were divided into two groups with a boundary of 20 mm,and further subdivided into four subgroups of <10mm,10～20mm,20～50mm,and ≥50mm,to verify the detection rates for lesions of different sizes for the 2.5D and 3D models.Mc Nemar’s test was used to compare the detection sensitivities among different methods.Experiment 2 We further explored the efficacy of multi-task CNNs framework together with multi-sequence MRI scans in the diagnosis of HCC.A total of 468 lesions from 406 patients at high risk for HCC(including 285 HCC,47 non-HCC malignancies and 136 benign lesions)were included in the study.There were two groups of classification tasks in this study.The first task was to classify all the tumors into two categories of HCC and non-HCC.The second task was to divide the tumors into three categories: HCC,non-HCC,benign lesions.The CNN outputs were compared with the results of three experts and three novices which was diagnosed based on 2018 LI-RADS category.For the two-way classification task,Mc Nemar’s test was used to compare the differences of sensitivity and specificity between readers,as well as between the CNN model and six readers.And for three-way classification task,the accuracy of CNN model and six readers were calculated respectively.Experiment 3 Pathologically confirmed HCC patients from two independent institutions were included in this study to explore the value of two-stage CNN framework in MVI prediction.Patients from Institution I were randomly divided into training(n=260)and internal validation cohort(n=84)and patients from Institution II served as an external validation cohort(n=81).Clinical model was established by clinical variables and radiological characteristics.Based on pre-contrast and three contrast-enhanced sequences,four monophasic models were compared to select two optimal phases for fusion model.And integrating clinical and fusion model to develop a combined model.De Long’s test was used to compare the AUCs of the four kinds of models to determine their predictive performances.ResultsExperiment 1 The results of automatic detection of FLLs based on multi-phasic CT images and multi-dimensional CNN frameworks showed that,portal venous phase outperformed arterial phase,and a combination of the two phases further improved the detection rate over single phase.The 3D model was superior to the 2.5D model for detecting benign lesions(0.896 vs 0.807,P<0.001),malignant lesions(0.940 vs 0.918,P =0.013)and all lesions(0.902 vs 0.832,P <0.001)regardless of size division.Particularly,the 3D model showed higher sensitivity than the 2.5D model in detecting lesions smaller than 20 mm,for benign lesions(0.871 vs 0.760,P <0.001),malignant lesions(0.846 vs 0.747,P <0.001)and all lesions(0.868 vs 0.759,P <0.001).While for lesions larger than 20 mm,both the 3D and the 2.5D models achieved excellent detection performance.Experiment 2 For two-way classification task,CNN model had higher accuracy than six readers(0.904 vs 0.809～0.904),showed significantly higher sensitivity than three novices(0.982 vs 0.821～0.893,P < 0.001)and exhibited equivalent specificity to six readers(0.789 vs 0.789～0.895,P > 0.05).The results of three-way classification task showed that the diagnostic accuracy of CNN model was higher than novices(0.894 vs 0.798～0.840)and similar to experts(0.894 vs 0.840～0.904).For both tasks,compared with group larger than 20 mm,the diagnostic accuracy and sensitivity for tumors smaller than 20 mm among six readers were significantly reduced,but the diagnostic performance of the CNN model was hardly affected by tumor size.Experiment 3 In this study,seven MVI status prediction models were established,including one clinical model,four monophasic models,one fusion model,and a combined model.Clinical models were composed of three independent MVI predictors,such as AFP≥20 ng/ml,enhancing capsule and corona enhancement,and its AUCs in the training,internal and external validation cohort were 0.678(0.609,0.746),0.698(0.574,0.822)and 0.691(0.564,0.817),respectively.Fusion model based on arterial and portal venous phase showed significantly higher predictive performance than clinical model,and AUCs of three cohorts were 0.952(0.928,0.976),0.833(0.726,0.941)and 0.805(0.683,0.927),which was equivalent to the combined model(P > 0.05).ConclusionsThe CNN models constructed in this paper showed good performance in multi-modality image analysis in FLLs detection,HCC diagnosis and MVI prediction.The results were summarized as follows.(1)The proposed 3D CNN detection model with multi-phasic CT images was demonstrated to generalize well to lesions of diverse types and various sizes.And 3D CNN framework showed an enhanced capability over the 2.5D in the detection of FLLs,particularly small lesions.(2)The diagnostic performance of CNN model for two-way and three-way classification task was significantly higher than novices,and even reaching the expert level.Our CNN model could reduce missed diagnosis of small lesions and facilitate early diagnosis and treatment,especially for novices.(3)The fusion model based on the arterial and portal venous phase performed well in the prediction of MVI,which was significantly higher than clinical model and comparable with combined model.Fusion model might be potentially helpful for MVI status prediction and accurate location in the clinical practice. |
PDF Full Text Request