Non-invasive Radiomics Model for Preoperative Prediction of Dual-phenotype Hepatocellular Carcinoma Based on CT Data

Non-invasive Radiomics Model for Preoperative Prediction of Dual-phenotype Hepatocellular Carcinoma Based on CT Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Non-invasive Radiomics Model for Preoperative Prediction of Dual-phenotype Hepatocellular Carcinoma Based on CT Data Lintao Chen, Yifan Wang, Jing Zhang, Hongbin Zhang, Ziqian Li, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6285893/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract BACKGROUND. Dual-phenotype hepatocellular carcinoma (DPHCC), a highly aggressive subtype, is difficult to diagnose preoperatively based on morphological characteristics. OBJECTIVE. We aimed to develop a radiomics model based on computed tomography (CT) data for non-invasive preoperative identification of DPHCC. METHODS. CT images and clinical data of 258 patients from Institution I were included in this study as a training set. Among them, 119 patients were diagnosed with DPHCC, and the rest 139 patients were treated as a control group. Radiomics features were extracted from regions of interest (ROI), and the features were selected by recursive feature elimination with cross-validation (RFECV). The logistic regression (LR) and random forest (RF) algorithm were used to develop four models to differentiate DPHCC and non-DPHCC, including the radiomics model, the clinical model, the radiologic model and the fused model. In addition, the effectiveness of the prediction models was evaluated by various indexes. CT images and clinical data of 58 patients from Institution II were included as an independent validation set. The models were evaluated on the independent validation set to assess the robustness of these models. RESULTS. Among these models, the radiomics model shows a balance of effectiveness and robustness. For the radiomics model, the mean area under the curve (AUC) of five-fold cross-validation on the training set reached 0.753, while the AUC on the independent validation set was 0.734. A radiomics signature containing 35 radiomics features was established for non-invasive prediction of dual-phenotype hepatocellular carcinoma. The fused model exhibits the best performance, with a training set AUC of 0.885 and an independent validation set AUC of 0.763. CONCLUSION. CT radiomics features of the tumor correlate significantly with the expression status of the bile duct phenotype in HCC. The radiomics feature, AFP level, ill-defined margin, pseudo-capsule and intra-tumoral necrosis are the essential features of the prediction model, potentially influencing the development of individualized treatment strategies for HCC. CLINICAL IMPACT. The radiomics model based on computed tomography (CT) data contribute to non-invasive preoperative identification of DPHCC. dual-phenotype hepatocellular carcinoma radiomics CT hepatocellular carcinoma Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Highlights Key Finding The mean area under the curve (AUC) of the radiomics, clinical, and radiologic models were 0.753, 0.779, 0.758, respectively. The radiomics model reported the best robustness at a mean AUC of 0.734 on the independent validation set. The fused model exhibits the best performance, with a training set AUC of 0.885 and an independent validation set AUC of 0.763. Feature Importance The radiomics feature, AFP level, ill-defined margin, pseudo-capsule and intra-tumoral necrosis are the essential features of the prediction model. Background Liver cancer is the sixth most common malignancy and the third most lethal cancer worldwide [ 1 , 2 ]. Hepatocellular carcinoma (HCC), the primary type of liver cancer, accounts for approximately 75–85% of cases, and the untreated were associated with poor prognoses [ 1 , 3 ]. Given high intrinsic heterogeneity, HCC can present as different pathological subtypes with different prognoses [ 4 ]. In recent years, dual-phenotype hepatocellular carcinoma (DPHCC) has gained considerable attention as a novel subtype of HCC with a higher postoperative recurrence rate and lower survival rate [ 5 – 7 ]. Unlike combined hepatocellular-cholangiocarcinoma (cHCC-CC), DPHCC has typical HCC histologic features and expresses both hepatocellular and cholangiocellular phenotypes [ 8 ]. Morphology-based liver cancer diagnosis and treatment guidelines are relatively inadequate for DPHCC. Therefore, there is an urgent clinical need for a definitive preoperative diagnosis of DPHCC to devise an individualized treatment strategy to improve prognosis. Currently, morphological features are incompetent to predict pathological subtypes of HCC, whereas the radiomics method offers potential for pathological feature prediction [ 9 ]. Contrast-enhanced CT and MRI imaging play an essential role in evaluating HCC, and even a clinical diagnosis could be made if the tumor exhibits typical HCC radiologic characteristics [ 10 , 11 ]. Prior studies have identified radiologic characteristics such as irregular margins, rim enhancement, and hypovascularity as predictive indicators for DPHCC [ 12 – 14 ]. However, the number of included patients was relatively small, and the diagnostic value of these non-quantitative features varied among the studies. In contrast to morphological features, thousands of quantitative radiomics features can be used to develop models for segmentation, diagnosis, staging, and prognostic prediction of HCC [ 15 ]. Radiomics models have also been used to predict HCC aggressive phenotypes and immunophenotypes [ 16 ]. Several investigators have developed radiomics models for diagnosing DPHCC using MRI data, yielding good preoperative predictive efficacy [ 17 – 21 ]. These studies demonstrated the potential of radiomics as a non-invasive tool for classifying HCC pathological subtypes by decoding MRI data. To the best of our knowledge, there is still a lack of a radiomics model based on CT data for the preoperative prediction of DPHCC. Therefore, we developed a radiomics model using CT data combined with clinical characteristics for the preoperative diagnosis of DPHCC. Methods Ethical Information and Data Collection We obtained ethical approval from the institutional review boards of the First Affiliated Hospital, Zhejiang University School of Medicine (Institution I) and the First Affiliated Hospital of Zhengzhou University (Institution II). The requirement for informed consent is waived. Between June 2011 and December 2016, we retrospectively enrolled 258 HCC patients at Institution I as a training set, of which 119 had DPHCC and 139 had non-DPHCC. The inclusion criteria for the study were surgically resected tumors with pathological histological and immunohistochemical confirmation of DPHCC, while non-DPHCC served as the control group. A hepatic contrast-enhanced CT scan was performed within one month prior to surgery, and complete clinical data, CT data, and pathologic specimens were available to analyze HCC cases. We excluded patients if their pathologic results revealed a non-HCC tumor, had received anti-tumor therapy prior to surgery, or if their CT image was insufficient for radiomics analysis. We also retrospectively enrolled 58 HCC patients at Institution II following the same criteria, of which 32 had DPHCC and 26 had non-DPHCC. During model evaluation, the dataset from Institution II served as the independent validation set. The case recruitment process for this study is shown in Fig. 1 . CT Scan Parameters For Institution I, all patients underwent preoperative contrast-enhanced CT scans using imaging equipment, including a 16-slice CT (Aquilion 16, Toshiba Medical Systems, Japan) and a 256-slice CT (Brilliance iCT, Philips Medical Systems, The Netherlands). For Institution II, all patients underwent preoperative contrast-enhanced CT scans using imaging equipment, including a 256-slice CT (Revolution CT, GE Healthcare, United States) and a 320-MDCT (Aquilion ONE, Otawara, Japan). The scanning protocol utilized the following parameters: tube voltage of 120–125 kV, tube current of 50–500 mAs, matrix of 512×512, a pitch of 0.95 mm, a field of view of 500×500 mm, slice thickness ranging from 2–5 mm, and reconstruction interval of 2–5 mm. A dynamic three-phase contrast-enhanced CT scan of the entire upper abdomen was performed for patients in both groups. The contrast agent was injected through the forearm vein, and the arterial, portal, and delayed phases were scanned at 20–35 seconds, 55–70 seconds, and (or) 120–140 seconds after injection, respectively. The injection volume was 1.5 mL/kg at a high-pressure injector velocity of 2.5-4.0 mL/s. Clinical and Radiologic Characteristics We collected and analyzed the clinical and radiologic characteristics of the DPHCC and non-DPHCC groups. The clinical data included gender, age, chronic hepatitis B virus (HBV) infection, cirrhosis, clinical symptom, and preoperative serum AFP. Meanwhile, the radiologic characteristics included location, size, morphology, margin, intra-tumoral fat, intra-tumoral hemorrhage, intra-tumoral necrosis, arterial phase rim enhancement, pseudo-capsule, non-peripheral washout, progressive enhancement, intrahepatic metastasis, and lymphadenopathy. The CT data were reviewed by two experienced junior radiologists specializing in abdominal tumors, who did not know the final pathological findings. In case of disagreement, a senior radiologist was consulted for final evaluation. Radiomics Analysis In this radiomics study, the portal phase CT images of HCC were manually segmented using ITK-SNAP (v3.6.0) software. The section with the largest tumor area was segmented twice by two radiologists and reviewed by a senior radiologist to ensure accuracy. Radiomics Feature Extraction and Selection Before feature extraction, preprocessing was conducted using isotropic resampling and uniform quantization methods. Three types of radiomics features were extracted: first-order features, texture features, and high-order filtering features. The high-order filtering features included Haar wavelet features, sym8 wavelet features, and local binary pattern (LBP) features. Intraclass correlation coefficients (ICCs) were calculated between the two feature sets. Features with low repeatability were identified and eliminated. The variance threshold selection algorithm was then used for other feature selection, and new features were generated using the symbolic regression method. The most valuable feature set was obtained through a recursive feature elimination with a cross-validation (RFECV) process. Radiomics Model Construction A radiomics model was developed to distinguish between dual-phenotype HCC (DPHCC) and non-DPHCC using a logistic regression (LR) classifier. In preliminary experiments, we also tried support vector machine (SVM) and random forest (RF) classifier. Among these models, logistic regression has better performance and interpretability. The five-fold cross-validation method with stratified sampling was employed to reduce overfitting and maximize the use of training set. Clinical and radiologic models were constructed using RF and feature selection algorithms. The critical clinical and radiologic characteristics were selected using similar methods to those used in constructing the radiomics model. The prediction probability obtained from the radiomics model was used as the radiomics score and incorporated into the fused feature set, which included clinical and radiologic characteristics. The RFECV method was also used for feature selection, and the fused model was constructed using the random forest algorithm. Model Evaluation The performance of the four models was evaluated using quantitative indexes such as the area under the curve (AUC), accuracy, sensitivity, and specificity. Receiver operating characteristic (ROC), calibration, and decision curves were also drawn to compare the performance of the four models visually. The study flowchart is presented in Fig. 2 . The performance of the models was also evaluated on an independent validation set, in order to compare the stability of different features. Survival Analysis Our study employed the Kaplan-Meier method for survival analysis to examine the prognostic value of the models. The follow-up data of the patients is available, including the operation time and the recurrence time. Patients were assigned to 'DPHCC' group and 'non-DPHCC' group. They were simultaneously assigned to 'predicted DPHCC' group and 'predicted non-DPHCC' group based on the prediction results of the radiomics model. Statistical Analysis During data processing and statistical analysis, we used specific software, including MATLAB (R2021a, MathWorks, Natick, MA, USA), Python 3.7.13 ( https://www.python.org/ ), and R 4.2.3 ( http://www.Rproject.org ). We extracted radiomics features using MATLAB libraries such as 'Wavelet,' 'Communication Toolbox,' and 'Radiomics.' We analyzed the data using SPSS 26 (IBM Corp, Chicago, USA). Continuous variables were presented as 'mean ± standard deviation,' while categorical variables were presented as values and corresponding proportions. We used the chi-square test to analyze differences between groups for unordered categorical variables and the t-test for continuous variables with a normal distribution. Variables with p < .05 were considered statistically significant. Results Clinical and Radiologic Characteristics The DPHCC group consisted of 119 patients (mean age: 52.0 ± 10.4 years) with an average tumor size of 4.6 ± 2.8 cm. The non-DPHCC group comprised 139 patients (mean age: 59.0 ± 11.4 years) with an average tumor size of 5.6 ± 3.0 cm. The model-building process included five clinical characteristics: male gender, chronic HBV infection, cirrhosis, clinical symptoms, and AFP. Twelve radiologic characteristics were considered during the model construction process, including right lobe, irregular shape, ill-defined, intra-tumoral fat, intra-tumoral hemorrhage, intra-tumoral necrosis, arterial phase rim enhancement, pseudo-capsule, non-peripheral washout, progressive enhancement, intrahepatic metastasis, and lymphadenopathy. The clinical and radiologic data are detailed in Table S1 and Table S2, respectively. Radiomics Feature Extraction and Selection A total of 1067 radiomics features were extracted, including 7 first-order features, 22 GLCM texture features, 13 GLRLM texture features, 13 GLSZM texture features, 5 NGTDM texture features, 424 Haar wavelet features, 424 sym8 wavelet features, and 159 LBP features. After applying an ICC threshold of 0.9, 816 radiomics features with high repeatability were retained. Using a variance threshold of 0.05, more than half of the remaining features were discarded, resulting in 339 radiomics features. In addition, 30 new features were generated through symbolic regression, bringing 369 features to the step of RFECV. The final feature set consisted of 35 key features. Model Construction The radiomics model was constructed using LR and five-fold cross-validation with stratified sampling based on the 35 most valuable radiomics features. The best parameters for the logistic regression algorithm of the radiomics model was: 'C': 12.915. The formula for radiomics signature is detailed in Results S1. The clinical model was constructed using RF and five-fold cross-validation based on all five clinical characteristics. The best parameters of the clinical model were: 'criterion': 'gini'; 'max_ depth': 2; 'max_features': 'auto'; 'min_samples_leaf': 20; 'min_samples_split': 2; ' n_estimators': 50. The radiologic model was constructed using RF based on the four most valuable features selected by RFECV: ill-defined, intra-tumoral hemorrhage, intra-tumoral necrosis and pseudo-capsule. The best parameters of the radiologic model were: 'criterion': 'entropy'; 'max_depth': 3; 'max_features': 'auto'; 'min_samples_leaf': 4; 'min_samples_split':2; 'n_estimators': 50. The fused model set consisted of 5 clinical characteristics, 12 radiologic characteristics, and the radiomics score. After RFECV, a key feature set with 14 features was obtained, including male gender, chronic HBV infection, cirrhosis, clinical symptoms, AFP, right lobe, irregular shape, ill-defined, intra-tumoral hemorrhage, intra-tumoral necrosis, pseudo-capsule, non-peripheral washout, intrahepatic metastasis and the radiomics score. The best parameters of the fused model were: 'criterion': 'gini'; 'max_depth': 10; 'max_features': 'sqrt'; 'min_samples_leaf': 1; 'min_samples_split': 5; 'n_estimators': 50. Feature importance of the fused model are shown in Fig. 3 . Model Evaluation The CT radiomics model was constructed using the LR algorithm and five-fold cross-validation, resulting in a mean AUC of 0.753, with an accuracy of 65.2%, a sensitivity of 67.2%, and a specificity of 63.4%. The clinical model had a mean AUC of 0.779, while the radiologic model had a mean AUC of 0.758. The fused model demonstrated the best performance, outperforming the other models, with a mean AUC of 0.885, an accuracy of 81.4%, a sensitivity of 76.4%, and a specificity of 85.7%. Table 1 shows the detailed quantitative indexes of the four models. Figure 4 displays the ROC curves that compare the performance of the four models. As shown in Fig. 5 , the calibration curves indicate that the predicted values of the four models fit the actual values well. The decision curves of the four models are presented in Fig. 6 , indicating that the fused model outperforms the other models. Additionally, the correlation heatmap of the fused feature set is illustrated in Fig. 7 . Table 1 Quantitative Indices of the Four Models for Predicting DPHCC Model Training Set Independent Validation Set AUC ACC SENS SPEC AUC Radiomics Model 0.753 0.652 0.672 0.634 0.734 Clinical Model 0.779 0.717 0.765 0.676 0.600 Radiologic Model 0.758 0.702 0.664 0.734 0.621 Fused Model 0.885 0.814 0.764 0.857 0.763 Note—AUC = area under the curve, ACC = accuracy, SENS = sensitivity, SPEC = specificity. Independent Validation In order to exclude the influence of scanning devices and scanning parameters on the stability of radiomics features, we used the radiomics model to make predictions on the independent validation set from Institution II, and obtained an AUC of 0.734, indicating that the radiomics model has reliable stability. As for the clinical model and radiologic model, the AUC on the independent validation set are 0.600 and 0.621, respectively, indicating that their stability is not as good as that of the radiomics model. The fused model achieved an AUC of 0.763 on the independent validation set, slightly higher than the AUC of the radiomics model on the independent validation set, but lower than its own training set AUC. This indicates that the addition of clinical and radiologic features improves the performance of the model, but at the same time, it also leads to unavoidable overfitting of the fused model. Survival Analysis As shown in Fig. 8 , patients from the DPHCC and non-DPHCC groups showed significant differences (p < 0.05) in the postoperative recurrence-free survival time, which reflected the value of DPHCC classification in predicting prognosis. At the same time, the survival analysis results of the predicted DPHCC and predicted non-DPHCC groups also confirmed the potential predictive value of the radiomics model classification results for the recurrence-free survival time of patients. Discussion In this study, we used a CT dataset to construct an artificial intelligence algorithm for preoperative identification of DPHCC. Based on the random forest algorithm, the fused model utilized radiomics features, radiologic characteristics, and clinical characteristics. Compared to models constructed with single-class features, the fused model provided a non-invasive, highly effective, and individualized diagnostic tool for diagnosing and managing liver cancer. This preoperative diagnostic tool for DPHCC provides valuable information for developing clinical management strategies. To facilitate the use of our DPHCC prediction model, we have made our algorithm publicly available on GitHub ( https://github.com/YfWangZJU/DPHCC.git ). Previous studies have highlighted the importance of wavelet features in constructing radiomics models. In two MRI-based radiomics models for diagnosing DPHCC, wavelet features accounted for 15 out of 21 and 6 out of 14 of the total features [ 18 , 21 ]. Based on these findings, we included Haar and sym8 wavelet bases in our study to further reflect the significance of wavelet features. In addition, Wang et al. [ 19 ] reported the significance of texture features in predicting the cytokeratin 19 status of HCC, emphasizing the need to combine different features for good prediction efficiency. Unlike most previous studies, we used the symbolic transformation method to generate new synthetic features, balancing a limited number of features with maximizing the use of different features at an early stage of feature selection. Our radiomics model consists of 35 key features, of which 3 features were generated through symbolic regression, which has been shown to be more effective and robust than the original features in our results. Two were local binary pattern features used to describe image texture. The remaining features came from wavelet transform, highlighting the importance of wavelet features, which was consistent with existing research findings. Our research has found that high-order filtering features have significant value in the differential diagnosis of DPHCC In recent years, there has been an ongoing discussion about the effectiveness of 2D versus 3D segmentation in radiomics studies. In most studies of HCC, 2D and 3D radiomics features have been extracted jointly by 3D segmentation [ 18 , 19 , 21 ]. A previous study also compared the prediction performance of radiomics models based on 2D and 3D segmentation, with the latter showing superior performance [ 22 ]. Consistent with the previous results, our radiomics model based on 3D segmentation yielded a mean AUC of 0.850 (Figure S1 ) and an accuracy of 0.70.1%, which was better than the radiomics model based on 2D segmentation. Meanwhile, the 3D-fused model achieved a mean AUC of 0.875 (Figure S2) and an accuracy of 0.76.8%. Details of the 3D radiomics model and the 3D-fused model are described in Results S2. Although 3D radiomics features provide more information about intra-tumoral heterogeneity, the choice between 2D and 3D segmentation is a trade-off between time cost, interpretability, and predictive performance. In our opinion, the 2D-fused model balances these factors well in this study, as the performance of the 3D-fused model was not significantly superior to that of the 2D-fused model. In our study, as shown in Fig. 3 , we observed that specific radiologic characteristics, such as an ill-defined margin, intra-tumoral hemorrhage, intra-tumor necrosis and pseudo-capsule, could improve the efficacy of the radiomics model in predicting DPHCC. Previous research has indicated that particular morphological features can be imaging biomarkers of tumor invasion [ 23 , 24 ]. In the beginning, Chung et al. studied a cohort of hypovascular HCC cases on CT with a poor prognosis associated with CK19 expression [ 12 ]. Additionally, an MRI investigation found that irregular margins and rim enhancement were predictors for diagnosing CK19-positive HCC [ 13 ]. However, our radiologic findings differed from previous research, possibly due to variations in the DPHCC population and imaging modalities. While two studies have demonstrated that radiologic characteristics improve predictive efficacy for DPHCC [ 19 , 20 ], others did not incorporate radiologic characteristics [ 18 , 21 ]. As shown in Fig. 7 , our study revealed a weak correlation between radiologic characteristics, radiomics features, and clinical characteristics, supporting the independent predictive role of radiologic characteristics in the model. Thus, we recommend carefully defining radiologic characteristics and judiciously incorporating them to construct effective DPHCC prediction models. At the same time, we find that the radiomics model performs better than the clinical model and the radiologic model on the independent validation set, indicating that radiomics features have stronger robustness and predictive value. This confirms that computer vision algorithms can extract high-dimensional features that are difficult for the human eye to detect for differential diagnosis of DPHCC. In this study, preoperative AFP levels were significantly higher in the DPHCC group than in the non-DPHCC group and made up a crucial predictive feature in the model. A high AFP level is widely recognized as an essential biomarker for diagnosing HCC in clinical liver cancer guidelines. Subsequent studies have also shown that high preoperative AFP levels were critical prognostic markers, predicting early recurrence and poor overall survival in patients with HCC after surgery [ 25 – 27 ]. In two recent independent studies of DPHCC using MRI data, AFP was also included in constructing radiomics models to improve its preoperative predictive power [ 19 , 20 ]. However, two other studies also yielded opposite results, with no significant difference in AFP positivity rates between the DPHCC and non-DPHCC groups. Therefore, AFP were not included in their prediction models [ 18 , 21 ]. These discrepant findings might be related to the studies' relatively small number of DPHCC cases and insufficient MRI data. In contrast, our findings supported that AFP levels contributed to the preoperative prediction of DPHCC and were an essential clinical characteristic in building a predictive model. Compared to traditional radiomics research, we adopted a multicenter study approach, which reduced the risk of overfitting in the model and improved the reliability of the results. However, our retrospective study still has some limitations. Firstly, data protection restrictions and difficulties in obtaining the bile duct immunophenotype for HCC made it challenging to conduct large-scale HCC radiomics studies. Secondly, our random forest-based algorithm for the DPHCC prediction model did not provide a nomogram for clinical use, which increased the difficulty of clinical practice of AI models. Nevertheless, the algorithm code and related data for predicting DPHCC using CT data were available to interested scholars. Thirdly, although including specific radiologic characteristics in the DPHCC model improved the predictive efficacy, the reproducibility of these qualitative characteristics may be a matter of concern. In conclusion, we used logistic regression algorithm to train a radiomics model with a balance of effectiveness and robustness. We also developed a random forest model incorporating radiomics features, clinical characteristics and radiologic characteristics for the preoperative prediction of DPHCC. This non-invasive DPHCC prediction model is an imaging marker that can potentially be used to develop individualized treatment strategies for liver cancer. Declarations Author Contribution CLT and WYF collected and collated the patient data and wrote the first draft. CLT, ZJ and LZQ conducted a data collection of patients. WYF analyzed the possible correlation between the radiomics features of colorectal cancer and machine learning. ZHB, SFY and DWT participated in the delineation of ROl. DY and LWJ put forward guiding opinions on the whole article and revised the whole paper. All authors reviewed the manuscript. References Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021; 71:209–249 Llovet JM, Kelley RK, Villanueva A, et al. Hepatocellular carcinoma. Nat Rev Dis Primers 2021; 7:6 Kulik L, El-Serag HB. Epidemiology and Management of Hepatocellular Carcinoma. Gastroenterology 2019; 156:477–491.e1 Li L, Wang H. Heterogeneity of liver cancer and personalized therapy. Cancer Lett 2016; 379:191–7 Lu XY, Xi T, Lau WY, et al. Hepatocellular carcinoma expressing cholangiocyte phenotype is a novel subtype with highly aggressive behavior. Ann Surg Oncol 2011; 18:2210–7 Zhuo JY, Lu D, Tan WY, et al. CK19-positive Hepatocellular Carcinoma is a Characteristic Subtype. J Cancer 2020; 11:5069–5077 Chiba T, Kamiya A, Yokosuka O, et al. Cancer stem cells in hepatocellular carcinoma: Recent progress and perspective. Cancer Lett 2009; 286:145–53 Cong WM, Bu H, Chen J, et al. Guideline Committee. Practice guidelines for the pathological diagnosis of primary liver cancer: 2015 update. World J Gastroenterol 2016; 22:9279–9287. Feng B, Ma XH, Wang S, et al. Application of artificial intelligence in preoperative imaging of hepatocellular carcinoma: Current status and future perspectives. World J Gastroenterol 2021; 27:5341–5350 Chernyak V, Fowler KJ, Kamaya A, et al. Liver Imaging Reporting and Data System (LI-RADS) Version 2018: Imaging of Hepatocellular Carcinoma in At-Risk Patients. Radiology 2018; 289:816–830 Zhou J, Sun H, Wang Z, et al. Guidelines for the Diagnosis and Treatment of Hepatocellular Carcinoma (2019 Edition). Liver Cancer 2020; 9:682–720 Chung GE, Lee JH, Yoon JH, et al. Prognostic implications of tumor vascularity and its relationship to cytokeratin 19 expression in patients with hepatocellular carcinoma. Abdom Imaging 2012; 37:439–46 Choi SY, Kim SH, Park CK, et al. Imaging Features of Gadoxetic Acid-enhanced and Diffusion-weighted MR Imaging for Identifying Cytokeratin 19-positive Hepatocellular Carcinoma: A Retrospective Observational Study. Radiology 2018; 286:897–908 Hu XX, Wang WT, Yang L, et al. MR features based on LI-RADS identify cytokeratin 19 status of hepatocellular carcinomas. Eur J Radiol 2019; 113:7–14 Lewis S, Hectors S, Taouli B. Radiomics of hepatocellular carcinoma. Abdom Radiol (NY) 2021; 46:111–123 Miranda J, Horvat N, Fonseca GM, et al. Current status and future perspectives of radiomics in hepatocellular carcinoma. World J Gastroenterol 2023; 29:43–60 Wang HQ, Yang C, Zeng MS, et al. Magnetic resonance texture analysis for the identification of cytokeratin 19-positive hepatocellular carcinoma. Eur J Radiol 2019; 117:164–170 Huang X, Long L, Wei J, et al. Radiomics for diagnosis of dual-phenotype hepatocellular carcinoma using Gd-EOB-DTPA-enhanced MRI and patient prognosis. J Cancer Res Clin Oncol 2019; 145:2995–3003 Wang W, Gu D, Wei J, et al. A radiomics-based biomarker for cytokeratin 19 status of hepatocellular carcinoma with gadoxetic acid-enhanced MRI. Eur Radiol 2020; 30:3004–3014 Chen Y, Chen J, Zhang Y, et al. Preoperative Prediction of Cytokeratin 19 Expression for Hepatocellular Carcinoma with Deep Learning Radiomics Based on Gadoxetic Acid-Enhanced Magnetic Resonance Imaging. J Hepatocell Carcinoma 2021; 8:795–808 Wu Q, Yu YX, Zhang T, et al. Preoperative Diagnosis of Dual-Phenotype Hepatocellular Carcinoma Using Enhanced MRI Radiomics Models. J Magn Reson Imaging 2023; 57:1185–1196 Xu L, Yang P, Yen EA, Wan Y, et al. A multi-organ cancer study of the classification performance using 2D and 3D image features in radiomics analysis. Phys Med Biol 2019; 64:215009 Cho ES, Choi JY. MRI features of hepatocellular carcinoma related to biologic behavior. Korean J Radiol 2015 May-Jun; 16:449 – 64 Yoneda N, Matsui O, Kobayashi S, et al. Current status of imaging biomarkers predicting the biological nature of hepatocellular carcinoma. Jpn J Radiol 2019; 37:191–208 Ma WJ, Wang HY, Teng LS. Correlation analysis of preoperative serum alpha-fetoprotein (AFP) level and prognosis of hepatocellular carcinoma (HCC) after hepatectomy. World J Surg Oncol 2013; 11:212 Chan MY, She WH, Dai WC, et al. Prognostic value of preoperative alpha-fetoprotein (AFP) level in patients receiving curative hepatectomy- an analysis of 1,182 patients in Hong Kong. Transl Gastroenterol Hepatol 2019; 4:52 Zhang S, Xu L, Dai F, et al. Construction of a predictive nomogram and bioinformatic investigation of the potential mechanism of postoperative early recurrence of hepatocellular carcinoma meeting the Milan criteria. Ann Transl Med 2022; 10:866. Additional Declarations No competing interests reported. Supplementary Files SupplementalMaterial.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6285893","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":435965123,"identity":"16e225c9-4402-4577-bd90-d714ffd03f34","order_by":0,"name":"Lintao Chen","email":"","orcid":"","institution":"First Affiliated Hospital Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Lintao","middleName":"","lastName":"Chen","suffix":""},{"id":435965124,"identity":"cbd52a5c-7d30-49d8-913f-1e396f67fdf8","order_by":1,"name":"Yifan Wang","email":"","orcid":"","institution":"College of Integrated Circuits, Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Yifan","middleName":"","lastName":"Wang","suffix":""},{"id":435965125,"identity":"7e378823-97a0-4d6e-8234-b6a5b1ed5742","order_by":2,"name":"Jing Zhang","email":"","orcid":"","institution":"First Affiliated Hospital Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Zhang","suffix":""},{"id":435965126,"identity":"cf6540f9-2084-4268-aed3-32c56eb5acc5","order_by":3,"name":"Hongbin Zhang","email":"","orcid":"","institution":"First Affiliated Hospital Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Hongbin","middleName":"","lastName":"Zhang","suffix":""},{"id":435965127,"identity":"e495764b-2c42-4206-b4db-6c8e97066e6f","order_by":4,"name":"Ziqian Li","email":"","orcid":"","institution":"First Affiliated Hospital Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Ziqian","middleName":"","lastName":"Li","suffix":""},{"id":435965128,"identity":"b0e95d9c-e786-42cd-8cc9-7348555695c5","order_by":5,"name":"Fangyu Sun","email":"","orcid":"","institution":"Xiaoshan Traditional Chinese Medical Hospital of Zhejiang Chinese Medical University","correspondingAuthor":false,"prefix":"","firstName":"Fangyu","middleName":"","lastName":"Sun","suffix":""},{"id":435965129,"identity":"dcdfa089-f513-4569-9075-95ae73b186da","order_by":6,"name":"Wenting Du","email":"","orcid":"","institution":"First Affiliated Hospital Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Wenting","middleName":"","lastName":"Du","suffix":""},{"id":435965130,"identity":"a89530bd-da06-43bd-ad96-cc0cc347861f","order_by":7,"name":"Peijie Lyu","email":"","orcid":"","institution":"First Affiliated Hospital of Zhengzhou University","correspondingAuthor":false,"prefix":"","firstName":"Peijie","middleName":"","lastName":"Lyu","suffix":""},{"id":435965131,"identity":"3d2508c3-f0ed-4eab-96fb-c2dc67a5e8c3","order_by":8,"name":"Mingliang Ying","email":"","orcid":"","institution":"Affiliated Jinhua Hospital, Zhejiang University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Mingliang","middleName":"","lastName":"Ying","suffix":""},{"id":435965132,"identity":"c3edbb34-fd38-400d-b399-ef615e0d61d3","order_by":9,"name":"Yong Ding","email":"","orcid":"","institution":"College of Integrated Circuits, Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Yong","middleName":"","lastName":"Ding","suffix":""},{"id":435965133,"identity":"c9f6dcb0-4188-423f-a550-65e56054b4e4","order_by":10,"name":"Wenjie Liang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAw0lEQVRIiWNgGAWjYBACAzBisGFgbABSPCRoSSNdy2EIjygt5uyHt0l83HHennlGAuODt20M8uaEtFj2pJVJzjxzm5lxRgKz4dw2BsOdDYQcdiDHTJq37TYbUAsbkMGQYHCAkJbzb8yk/7ad4wFqYf9NnJYbQFsY2w5IgGxhJkqL5YxnxZa9bckGjD0PmyXnnJMw3EBIizl/8sYbP9vs7A3bkw9+eFNmI0/QFjgwbABHpgSx6oFAngS1o2AUjIJRMMIAANGhPHJSE3k7AAAAAElFTkSuQmCC","orcid":"","institution":"First Affiliated Hospital Zhejiang University","correspondingAuthor":true,"prefix":"","firstName":"Wenjie","middleName":"","lastName":"Liang","suffix":""}],"badges":[],"createdAt":"2025-03-23 00:53:04","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6285893/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6285893/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":79758609,"identity":"f89bbfd2-708d-4a18-9072-31fadd0d746a","added_by":"auto","created_at":"2025-04-02 10:47:57","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":51778,"visible":true,"origin":"","legend":"\u003cp\u003eThe case recruitment process. Dual-phenotype hepatocellular carcinoma = DPHCC.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/5ce79fbd68dd321a7c49e42f.png"},{"id":79757812,"identity":"bf7f2b7e-d869-4426-9bfd-cf49144cbe4e","added_by":"auto","created_at":"2025-04-02 10:39:57","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":131241,"visible":true,"origin":"","legend":"\u003cp\u003eStudy flowchart. CT = computed tomography, ROI = region of interest, ROC = receiver operating characteristic.\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/0827c13f40f87670ce6878cd.jpg"},{"id":79759647,"identity":"50456e9b-35e9-4119-8ad7-482c7485c99b","added_by":"auto","created_at":"2025-04-02 10:55:57","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":77507,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance of the fused model.\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/1fc1fe9c224f894963308cf0.jpg"},{"id":79757839,"identity":"a44e88ef-cd51-4d54-9442-a398a5348399","added_by":"auto","created_at":"2025-04-02 10:39:59","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":112210,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curves that compare the performance of the four models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA, \u003c/strong\u003eROC curves of five-fold cross-validation of the radiomics model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB,\u003c/strong\u003eROC curves of five-fold cross-validation of the clinical model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eC,\u003c/strong\u003eROC curves of five-fold cross-validation of the radiologic model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eD,\u003c/strong\u003eROC curves of five-fold cross-validation of the fused model\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/d71f6e18041fe5efe6802a8e.jpg"},{"id":79757820,"identity":"2d9e8e0a-9d54-4b91-9d3e-6809b0902197","added_by":"auto","created_at":"2025-04-02 10:39:57","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":142824,"visible":true,"origin":"","legend":"\u003cp\u003eCalibration curves of the radiomics model, the clinical model, the radiologic model, and the fused model.\u003c/p\u003e","description":"","filename":"Picture5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/f4f80afee7b866317cffe0ab.jpg"},{"id":79757840,"identity":"b55630f4-347e-4926-b7e8-b8164da05245","added_by":"auto","created_at":"2025-04-02 10:39:59","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":162835,"visible":true,"origin":"","legend":"\u003cp\u003eDecision curves of the radiomics model, the clinical model, the radiologic model, and the fused model.\u003c/p\u003e","description":"","filename":"Picture6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/2d9e04002ec5cd097d5c515f.jpg"},{"id":79760168,"identity":"352a01c9-c5e4-43d3-84f4-dc7e9aa8c4d8","added_by":"auto","created_at":"2025-04-02 11:03:57","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":205583,"visible":true,"origin":"","legend":"\u003cp\u003eCorrelation heatmap of the fused feature set.\u003c/p\u003e","description":"","filename":"Picture7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/72c942923b03cbceb96fbedd.jpg"},{"id":79758617,"identity":"e42df505-e5e3-4644-b03a-7e225d6baf8f","added_by":"auto","created_at":"2025-04-02 10:47:57","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":131010,"visible":true,"origin":"","legend":"\u003cp\u003eSurvival curves of the DPHCC, non-DPHCC, predicted DPHCC and predicted non-DPHCC group.\u003c/p\u003e","description":"","filename":"Picture8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/ac35385b5a8e4752c3b794f1.jpg"},{"id":80499028,"identity":"248e8c20-4194-45c8-b673-2fb0edc15db1","added_by":"auto","created_at":"2025-04-14 02:31:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1794908,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/799d271e-d416-4d8d-8d0b-6bee8dfd29e4.pdf"},{"id":79757817,"identity":"dc17f11c-9953-4f24-aa76-b25b28993a05","added_by":"auto","created_at":"2025-04-02 10:39:57","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":253929,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-6285893/v1/d6a178b080aee6087d68ff70.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Non-invasive Radiomics Model for Preoperative Prediction of Dual-phenotype Hepatocellular Carcinoma Based on CT Data","fulltext":[{"header":"Highlights","content":"\u003cp\u003e\u003cstrong\u003eKey Finding\u003c/strong\u003e\u003c/p\u003e\n\u003cul class=\"decimal_type\" start=\"50\"\u003e\n \u003cli\u003eThe mean area under the curve (AUC) of the radiomics, clinical, and radiologic models were 0.753, 0.779, 0.758, respectively.\u003c/li\u003e\n \u003cli\u003eThe radiomics model reported the best robustness at a mean AUC of 0.734 on the independent validation set.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eThe fused model exhibits the best performance, with a training set AUC of 0.885 and an independent validation set AUC of 0.763.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003eFeature Importance\u003c/strong\u003e\u003c/p\u003e\n\u003cul start=\"50\"\u003e\n \u003cli\u003eThe radiomics feature, AFP level, ill-defined margin, pseudo-capsule and intra-tumoral necrosis are the essential features of the prediction model.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e"},{"header":"Background","content":"\u003cp\u003eLiver cancer is the sixth most common malignancy and the third most lethal cancer worldwide [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Hepatocellular carcinoma (HCC), the primary type of liver cancer, accounts for approximately 75\u0026ndash;85% of cases, and the untreated were associated with poor prognoses [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Given high intrinsic heterogeneity, HCC can present as different pathological subtypes with different prognoses [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. In recent years, dual-phenotype hepatocellular carcinoma (DPHCC) has gained considerable attention as a novel subtype of HCC with a higher postoperative recurrence rate and lower survival rate [\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Unlike combined hepatocellular-cholangiocarcinoma (cHCC-CC), DPHCC has typical HCC histologic features and expresses both hepatocellular and cholangiocellular phenotypes [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Morphology-based liver cancer diagnosis and treatment guidelines are relatively inadequate for DPHCC. Therefore, there is an urgent clinical need for a definitive preoperative diagnosis of DPHCC to devise an individualized treatment strategy to improve prognosis.\u003c/p\u003e \u003cp\u003eCurrently, morphological features are incompetent to predict pathological subtypes of HCC, whereas the radiomics method offers potential for pathological feature prediction [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Contrast-enhanced CT and MRI imaging play an essential role in evaluating HCC, and even a clinical diagnosis could be made if the tumor exhibits typical HCC radiologic characteristics [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Prior studies have identified radiologic characteristics such as irregular margins, rim enhancement, and hypovascularity as predictive indicators for DPHCC [\u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. However, the number of included patients was relatively small, and the diagnostic value of these non-quantitative features varied among the studies. In contrast to morphological features, thousands of quantitative radiomics features can be used to develop models for segmentation, diagnosis, staging, and prognostic prediction of HCC [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Radiomics models have also been used to predict HCC aggressive phenotypes and immunophenotypes [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Several investigators have developed radiomics models for diagnosing DPHCC using MRI data, yielding good preoperative predictive efficacy [\u003cspan additionalcitationids=\"CR18 CR19 CR20\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. These studies demonstrated the potential of radiomics as a non-invasive tool for classifying HCC pathological subtypes by decoding MRI data.\u003c/p\u003e \u003cp\u003eTo the best of our knowledge, there is still a lack of a radiomics model based on CT data for the preoperative prediction of DPHCC. Therefore, we developed a radiomics model using CT data combined with clinical characteristics for the preoperative diagnosis of DPHCC.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eEthical Information and Data Collection\u003c/h2\u003e \u003cp\u003e We obtained ethical approval from the institutional review boards of the First Affiliated Hospital, Zhejiang University School of Medicine (Institution I) and the First Affiliated Hospital of Zhengzhou University (Institution II). The requirement for informed consent is waived. Between June 2011 and December 2016, we retrospectively enrolled 258 HCC patients at Institution I as a training set, of which 119 had DPHCC and 139 had non-DPHCC. The inclusion criteria for the study were surgically resected tumors with pathological histological and immunohistochemical confirmation of DPHCC, while non-DPHCC served as the control group. A hepatic contrast-enhanced CT scan was performed within one month prior to surgery, and complete clinical data, CT data, and pathologic specimens were available to analyze HCC cases. We excluded patients if their pathologic results revealed a non-HCC tumor, had received anti-tumor therapy prior to surgery, or if their CT image was insufficient for radiomics analysis. We also retrospectively enrolled 58 HCC patients at Institution II following the same criteria, of which 32 had DPHCC and 26 had non-DPHCC. During model evaluation, the dataset from Institution II served as the independent validation set. The case recruitment process for this study is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCT Scan Parameters\u003c/h3\u003e\n\u003cp\u003eFor Institution I, all patients underwent preoperative contrast-enhanced CT scans using imaging equipment, including a 16-slice CT (Aquilion 16, Toshiba Medical Systems, Japan) and a 256-slice CT (Brilliance iCT, Philips Medical Systems, The Netherlands). For Institution II, all patients underwent preoperative contrast-enhanced CT scans using imaging equipment, including a 256-slice CT (Revolution CT, GE Healthcare, United States) and a 320-MDCT (Aquilion ONE, Otawara, Japan). The scanning protocol utilized the following parameters: tube voltage of 120\u0026ndash;125 kV, tube current of 50\u0026ndash;500 mAs, matrix of 512\u0026times;512, a pitch of 0.95 mm, a field of view of 500\u0026times;500 mm, slice thickness ranging from 2\u0026ndash;5 mm, and reconstruction interval of 2\u0026ndash;5 mm. A dynamic three-phase contrast-enhanced CT scan of the entire upper abdomen was performed for patients in both groups. The contrast agent was injected through the forearm vein, and the arterial, portal, and delayed phases were scanned at 20\u0026ndash;35 seconds, 55\u0026ndash;70 seconds, and (or) 120\u0026ndash;140 seconds after injection, respectively. The injection volume was 1.5 mL/kg at a high-pressure injector velocity of 2.5-4.0 mL/s.\u003c/p\u003e\n\u003ch3\u003eClinical and Radiologic Characteristics\u003c/h3\u003e\n\u003cp\u003eWe collected and analyzed the clinical and radiologic characteristics of the DPHCC and non-DPHCC groups. The clinical data included gender, age, chronic hepatitis B virus (HBV) infection, cirrhosis, clinical symptom, and preoperative serum AFP. Meanwhile, the radiologic characteristics included location, size, morphology, margin, intra-tumoral fat, intra-tumoral hemorrhage, intra-tumoral necrosis, arterial phase rim enhancement, pseudo-capsule, non-peripheral washout, progressive enhancement, intrahepatic metastasis, and lymphadenopathy. The CT data were reviewed by two experienced junior radiologists specializing in abdominal tumors, who did not know the final pathological findings. In case of disagreement, a senior radiologist was consulted for final evaluation.\u003c/p\u003e\n\u003ch3\u003eRadiomics Analysis\u003c/h3\u003e\n\u003cp\u003eIn this radiomics study, the portal phase CT images of HCC were manually segmented using ITK-SNAP (v3.6.0) software. The section with the largest tumor area was segmented twice by two radiologists and reviewed by a senior radiologist to ensure accuracy.\u003c/p\u003e\n\u003ch3\u003eRadiomics Feature Extraction and Selection\u003c/h3\u003e\n\u003cp\u003eBefore feature extraction, preprocessing was conducted using isotropic resampling and uniform quantization methods. Three types of radiomics features were extracted: first-order features, texture features, and high-order filtering features. The high-order filtering features included Haar wavelet features, sym8 wavelet features, and local binary pattern (LBP) features. Intraclass correlation coefficients (ICCs) were calculated between the two feature sets. Features with low repeatability were identified and eliminated. The variance threshold selection algorithm was then used for other feature selection, and new features were generated using the symbolic regression method. The most valuable feature set was obtained through a recursive feature elimination with a cross-validation (RFECV) process.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eRadiomics Model Construction\u003c/h2\u003e \u003cp\u003eA radiomics model was developed to distinguish between dual-phenotype HCC (DPHCC) and non-DPHCC using a logistic regression (LR) classifier. In preliminary experiments, we also tried support vector machine (SVM) and random forest (RF) classifier. Among these models, logistic regression has better performance and interpretability. The five-fold cross-validation method with stratified sampling was employed to reduce overfitting and maximize the use of training set.\u003c/p\u003e \u003cp\u003eClinical and radiologic models were constructed using RF and feature selection algorithms. The critical clinical and radiologic characteristics were selected using similar methods to those used in constructing the radiomics model.\u003c/p\u003e \u003cp\u003eThe prediction probability obtained from the radiomics model was used as the radiomics score and incorporated into the fused feature set, which included clinical and radiologic characteristics. The RFECV method was also used for feature selection, and the fused model was constructed using the random forest algorithm.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eModel Evaluation\u003c/h3\u003e\n\u003cp\u003eThe performance of the four models was evaluated using quantitative indexes such as the area under the curve (AUC), accuracy, sensitivity, and specificity. Receiver operating characteristic (ROC), calibration, and decision curves were also drawn to compare the performance of the four models visually. The study flowchart is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The performance of the models was also evaluated on an independent validation set, in order to compare the stability of different features.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eSurvival Analysis\u003c/h3\u003e\n\u003cp\u003eOur study employed the Kaplan-Meier method for survival analysis to examine the prognostic value of the models. The follow-up data of the patients is available, including the operation time and the recurrence time. Patients were assigned to 'DPHCC' group and 'non-DPHCC' group. They were simultaneously assigned to 'predicted DPHCC' group and 'predicted non-DPHCC' group based on the prediction results of the radiomics model.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eDuring data processing and statistical analysis, we used specific software, including MATLAB (R2021a, MathWorks, Natick, MA, USA), Python 3.7.13 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.python.org/\u003c/span\u003e\u003cspan address=\"https://www.python.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), and R 4.2.3 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.Rproject.org\u003c/span\u003e\u003cspan address=\"http://www.Rproject.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). We extracted radiomics features using MATLAB libraries such as 'Wavelet,' 'Communication Toolbox,' and 'Radiomics.' We analyzed the data using SPSS 26 (IBM Corp, Chicago, USA). Continuous variables were presented as 'mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation,' while categorical variables were presented as values and corresponding proportions. We used the chi-square test to analyze differences between groups for unordered categorical variables and the t-test for continuous variables with a normal distribution. Variables with p\u0026thinsp;\u0026lt;\u0026thinsp;.05 were considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eClinical and Radiologic Characteristics\u003c/h2\u003e \u003cp\u003eThe DPHCC group consisted of 119 patients (mean age: 52.0\u0026thinsp;\u0026plusmn;\u0026thinsp;10.4 years) with an average tumor size of 4.6\u0026thinsp;\u0026plusmn;\u0026thinsp;2.8 cm. The non-DPHCC group comprised 139 patients (mean age: 59.0\u0026thinsp;\u0026plusmn;\u0026thinsp;11.4 years) with an average tumor size of 5.6\u0026thinsp;\u0026plusmn;\u0026thinsp;3.0 cm. The model-building process included five clinical characteristics: male gender, chronic HBV infection, cirrhosis, clinical symptoms, and AFP. Twelve radiologic characteristics were considered during the model construction process, including right lobe, irregular shape, ill-defined, intra-tumoral fat, intra-tumoral hemorrhage, intra-tumoral necrosis, arterial phase rim enhancement, pseudo-capsule, non-peripheral washout, progressive enhancement, intrahepatic metastasis, and lymphadenopathy. The clinical and radiologic data are detailed in Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e and Table S2, respectively.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eRadiomics Feature Extraction and Selection\u003c/h2\u003e \u003cp\u003eA total of 1067 radiomics features were extracted, including 7 first-order features, 22 GLCM texture features, 13 GLRLM texture features, 13 GLSZM texture features, 5 NGTDM texture features, 424 Haar wavelet features, 424 sym8 wavelet features, and 159 LBP features. After applying an ICC threshold of 0.9, 816 radiomics features with high repeatability were retained. Using a variance threshold of 0.05, more than half of the remaining features were discarded, resulting in 339 radiomics features. In addition, 30 new features were generated through symbolic regression, bringing 369 features to the step of RFECV. The final feature set consisted of 35 key features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eModel Construction\u003c/h2\u003e \u003cp\u003eThe radiomics model was constructed using LR and five-fold cross-validation with stratified sampling based on the 35 most valuable radiomics features. The best parameters for the logistic regression algorithm of the radiomics model was: 'C': 12.915. The formula for radiomics signature is detailed in Results S1.\u003c/p\u003e \u003cp\u003eThe clinical model was constructed using RF and five-fold cross-validation based on all five clinical characteristics. The best parameters of the clinical model were: 'criterion': 'gini'; 'max_ depth': 2; 'max_features': 'auto'; 'min_samples_leaf': 20; 'min_samples_split': 2; ' n_estimators': 50.\u003c/p\u003e \u003cp\u003eThe radiologic model was constructed using RF based on the four most valuable features selected by RFECV: ill-defined, intra-tumoral hemorrhage, intra-tumoral necrosis and pseudo-capsule. The best parameters of the radiologic model were: 'criterion': 'entropy'; 'max_depth': 3; 'max_features': 'auto'; 'min_samples_leaf': 4; 'min_samples_split':2; 'n_estimators': 50.\u003c/p\u003e \u003cp\u003eThe fused model set consisted of 5 clinical characteristics, 12 radiologic characteristics, and the radiomics score. After RFECV, a key feature set with 14 features was obtained, including male gender, chronic HBV infection, cirrhosis, clinical symptoms, AFP, right lobe, irregular shape, ill-defined, intra-tumoral hemorrhage, intra-tumoral necrosis, pseudo-capsule, non-peripheral washout, intrahepatic metastasis and the radiomics score. The best parameters of the fused model were: 'criterion': 'gini'; 'max_depth': 10; 'max_features': 'sqrt'; 'min_samples_leaf': 1; 'min_samples_split': 5; 'n_estimators': 50. Feature importance of the fused model are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eModel Evaluation\u003c/h2\u003e \u003cp\u003eThe CT radiomics model was constructed using the LR algorithm and five-fold cross-validation, resulting in a mean AUC of 0.753, with an accuracy of 65.2%, a sensitivity of 67.2%, and a specificity of 63.4%. The clinical model had a mean AUC of 0.779, while the radiologic model had a mean AUC of 0.758. The fused model demonstrated the best performance, outperforming the other models, with a mean AUC of 0.885, an accuracy of 81.4%, a sensitivity of 76.4%, and a specificity of 85.7%. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the detailed quantitative indexes of the four models. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e displays the ROC curves that compare the performance of the four models. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, the calibration curves indicate that the predicted values of the four models fit the actual values well. The decision curves of the four models are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, indicating that the fused model outperforms the other models. Additionally, the correlation heatmap of the fused feature set is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eQuantitative Indices of the Four Models for Predicting DPHCC\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e \u003cp\u003eTraining Set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eIndependent Validation Set\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eACC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSENS\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSPEC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRadiomics Model\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.753\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.652\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.672\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.634\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.734\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eClinical Model\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.779\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.717\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.765\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.676\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.600\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRadiologic Model\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.758\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.702\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.664\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.734\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.621\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFused Model\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.885\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.814\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.764\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.857\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e0.763\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eNote\u0026mdash;AUC\u0026thinsp;=\u0026thinsp;area under the curve, ACC\u0026thinsp;=\u0026thinsp;accuracy, SENS\u0026thinsp;=\u0026thinsp;sensitivity, SPEC\u0026thinsp;=\u0026thinsp;specificity.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eIndependent Validation\u003c/h2\u003e \u003cp\u003eIn order to exclude the influence of scanning devices and scanning parameters on the stability of radiomics features, we used the radiomics model to make predictions on the independent validation set from Institution II, and obtained an AUC of 0.734, indicating that the radiomics model has reliable stability. As for the clinical model and radiologic model, the AUC on the independent validation set are 0.600 and 0.621, respectively, indicating that their stability is not as good as that of the radiomics model.\u003c/p\u003e \u003cp\u003eThe fused model achieved an AUC of 0.763 on the independent validation set, slightly higher than the AUC of the radiomics model on the independent validation set, but lower than its own training set AUC. This indicates that the addition of clinical and radiologic features improves the performance of the model, but at the same time, it also leads to unavoidable overfitting of the fused model.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eSurvival Analysis\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, patients from the DPHCC and non-DPHCC groups showed significant differences (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in the postoperative recurrence-free survival time, which reflected the value of DPHCC classification in predicting prognosis. At the same time, the survival analysis results of the predicted DPHCC and predicted non-DPHCC groups also confirmed the potential predictive value of the radiomics model classification results for the recurrence-free survival time of patients.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we used a CT dataset to construct an artificial intelligence algorithm for preoperative identification of DPHCC. Based on the random forest algorithm, the fused model utilized radiomics features, radiologic characteristics, and clinical characteristics. Compared to models constructed with single-class features, the fused model provided a non-invasive, highly effective, and individualized diagnostic tool for diagnosing and managing liver cancer. This preoperative diagnostic tool for DPHCC provides valuable information for developing clinical management strategies. To facilitate the use of our DPHCC prediction model, we have made our algorithm publicly available on GitHub (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/YfWangZJU/DPHCC.git\u003c/span\u003e\u003cspan address=\"https://github.com/YfWangZJU/DPHCC.git\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e \u003cp\u003ePrevious studies have highlighted the importance of wavelet features in constructing radiomics models. In two MRI-based radiomics models for diagnosing DPHCC, wavelet features accounted for 15 out of 21 and 6 out of 14 of the total features [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Based on these findings, we included Haar and sym8 wavelet bases in our study to further reflect the significance of wavelet features. In addition, Wang et al. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] reported the significance of texture features in predicting the cytokeratin 19 status of HCC, emphasizing the need to combine different features for good prediction efficiency. Unlike most previous studies, we used the symbolic transformation method to generate new synthetic features, balancing a limited number of features with maximizing the use of different features at an early stage of feature selection. Our radiomics model consists of 35 key features, of which 3 features were generated through symbolic regression, which has been shown to be more effective and robust than the original features in our results. Two were local binary pattern features used to describe image texture. The remaining features came from wavelet transform, highlighting the importance of wavelet features, which was consistent with existing research findings. Our research has found that high-order filtering features have significant value in the differential diagnosis of DPHCC\u003c/p\u003e \u003cp\u003eIn recent years, there has been an ongoing discussion about the effectiveness of 2D versus 3D segmentation in radiomics studies. In most studies of HCC, 2D and 3D radiomics features have been extracted jointly by 3D segmentation [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. A previous study also compared the prediction performance of radiomics models based on 2D and 3D segmentation, with the latter showing superior performance [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Consistent with the previous results, our radiomics model based on 3D segmentation yielded a mean AUC of 0.850 (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e) and an accuracy of 0.70.1%, which was better than the radiomics model based on 2D segmentation. Meanwhile, the 3D-fused model achieved a mean AUC of 0.875 (Figure S2) and an accuracy of 0.76.8%. Details of the 3D radiomics model and the 3D-fused model are described in Results S2. Although 3D radiomics features provide more information about intra-tumoral heterogeneity, the choice between 2D and 3D segmentation is a trade-off between time cost, interpretability, and predictive performance. In our opinion, the 2D-fused model balances these factors well in this study, as the performance of the 3D-fused model was not significantly superior to that of the 2D-fused model.\u003c/p\u003e \u003cp\u003eIn our study, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, we observed that specific radiologic characteristics, such as an ill-defined margin, intra-tumoral hemorrhage, intra-tumor necrosis and pseudo-capsule, could improve the efficacy of the radiomics model in predicting DPHCC. Previous research has indicated that particular morphological features can be imaging biomarkers of tumor invasion [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. In the beginning, Chung et al. studied a cohort of hypovascular HCC cases on CT with a poor prognosis associated with CK19 expression [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Additionally, an MRI investigation found that irregular margins and rim enhancement were predictors for diagnosing CK19-positive HCC [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. However, our radiologic findings differed from previous research, possibly due to variations in the DPHCC population and imaging modalities. While two studies have demonstrated that radiologic characteristics improve predictive efficacy for DPHCC [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], others did not incorporate radiologic characteristics [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e, our study revealed a weak correlation between radiologic characteristics, radiomics features, and clinical characteristics, supporting the independent predictive role of radiologic characteristics in the model. Thus, we recommend carefully defining radiologic characteristics and judiciously incorporating them to construct effective DPHCC prediction models. At the same time, we find that the radiomics model performs better than the clinical model and the radiologic model on the independent validation set, indicating that radiomics features have stronger robustness and predictive value. This confirms that computer vision algorithms can extract high-dimensional features that are difficult for the human eye to detect for differential diagnosis of DPHCC.\u003c/p\u003e \u003cp\u003eIn this study, preoperative AFP levels were significantly higher in the DPHCC group than in the non-DPHCC group and made up a crucial predictive feature in the model. A high AFP level is widely recognized as an essential biomarker for diagnosing HCC in clinical liver cancer guidelines. Subsequent studies have also shown that high preoperative AFP levels were critical prognostic markers, predicting early recurrence and poor overall survival in patients with HCC after surgery [\u003cspan additionalcitationids=\"CR26\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. In two recent independent studies of DPHCC using MRI data, AFP was also included in constructing radiomics models to improve its preoperative predictive power [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. However, two other studies also yielded opposite results, with no significant difference in AFP positivity rates between the DPHCC and non-DPHCC groups. Therefore, AFP were not included in their prediction models [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. These discrepant findings might be related to the studies' relatively small number of DPHCC cases and insufficient MRI data. In contrast, our findings supported that AFP levels contributed to the preoperative prediction of DPHCC and were an essential clinical characteristic in building a predictive model.\u003c/p\u003e \u003cp\u003eCompared to traditional radiomics research, we adopted a multicenter study approach, which reduced the risk of overfitting in the model and improved the reliability of the results. However, our retrospective study still has some limitations. Firstly, data protection restrictions and difficulties in obtaining the bile duct immunophenotype for HCC made it challenging to conduct large-scale HCC radiomics studies. Secondly, our random forest-based algorithm for the DPHCC prediction model did not provide a nomogram for clinical use, which increased the difficulty of clinical practice of AI models. Nevertheless, the algorithm code and related data for predicting DPHCC using CT data were available to interested scholars. Thirdly, although including specific radiologic characteristics in the DPHCC model improved the predictive efficacy, the reproducibility of these qualitative characteristics may be a matter of concern.\u003c/p\u003e \u003cp\u003eIn conclusion, we used logistic regression algorithm to train a radiomics model with a balance of effectiveness and robustness. We also developed a random forest model incorporating radiomics features, clinical characteristics and radiologic characteristics for the preoperative prediction of DPHCC. This non-invasive DPHCC prediction model is an imaging marker that can potentially be used to develop individualized treatment strategies for liver cancer.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eCLT and WYF collected and collated the patient data and wrote the first draft. CLT, ZJ and LZQ conducted a data collection of patients. WYF analyzed the possible correlation between the radiomics features of colorectal cancer and machine learning. ZHB, SFY and DWT participated in the delineation of ROl. DY and LWJ put forward guiding opinions on the whole article and revised the whole paper. All authors reviewed the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021; 71:209\u0026ndash;249\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLlovet JM, Kelley RK, Villanueva A, et al. Hepatocellular carcinoma. Nat Rev Dis Primers 2021; 7:6\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKulik L, El-Serag HB. Epidemiology and Management of Hepatocellular Carcinoma. Gastroenterology 2019; 156:477\u0026ndash;491.e1\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi L, Wang H. Heterogeneity of liver cancer and personalized therapy. Cancer Lett 2016; 379:191\u0026ndash;7\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu XY, Xi T, Lau WY, et al. Hepatocellular carcinoma expressing cholangiocyte phenotype is a novel subtype with highly aggressive behavior. Ann Surg Oncol 2011; 18:2210\u0026ndash;7\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhuo JY, Lu D, Tan WY, et al. CK19-positive Hepatocellular Carcinoma is a Characteristic Subtype. J Cancer 2020; 11:5069\u0026ndash;5077\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChiba T, Kamiya A, Yokosuka O, et al. Cancer stem cells in hepatocellular carcinoma: Recent progress and perspective. Cancer Lett 2009; 286:145\u0026ndash;53\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCong WM, Bu H, Chen J, et al. Guideline Committee. Practice guidelines for the pathological diagnosis of primary liver cancer: 2015 update. World J Gastroenterol 2016; 22:9279\u0026ndash;9287.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeng B, Ma XH, Wang S, et al. Application of artificial intelligence in preoperative imaging of hepatocellular carcinoma: Current status and future perspectives. World J Gastroenterol 2021; 27:5341\u0026ndash;5350\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChernyak V, Fowler KJ, Kamaya A, et al. Liver Imaging Reporting and Data System (LI-RADS) Version 2018: Imaging of Hepatocellular Carcinoma in At-Risk Patients. Radiology 2018; 289:816\u0026ndash;830\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou J, Sun H, Wang Z, et al. Guidelines for the Diagnosis and Treatment of Hepatocellular Carcinoma (2019 Edition). Liver Cancer 2020; 9:682\u0026ndash;720\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChung GE, Lee JH, Yoon JH, et al. Prognostic implications of tumor vascularity and its relationship to cytokeratin 19 expression in patients with hepatocellular carcinoma. Abdom Imaging 2012; 37:439\u0026ndash;46\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoi SY, Kim SH, Park CK, et al. Imaging Features of Gadoxetic Acid-enhanced and Diffusion-weighted MR Imaging for Identifying Cytokeratin 19-positive Hepatocellular Carcinoma: A Retrospective Observational Study. Radiology 2018; 286:897\u0026ndash;908\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu XX, Wang WT, Yang L, et al. MR features based on LI-RADS identify cytokeratin 19 status of hepatocellular carcinomas. Eur J Radiol 2019; 113:7\u0026ndash;14\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLewis S, Hectors S, Taouli B. Radiomics of hepatocellular carcinoma. Abdom Radiol (NY) 2021; 46:111\u0026ndash;123\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMiranda J, Horvat N, Fonseca GM, et al. Current status and future perspectives of radiomics in hepatocellular carcinoma. World J Gastroenterol 2023; 29:43\u0026ndash;60\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang HQ, Yang C, Zeng MS, et al. Magnetic resonance texture analysis for the identification of cytokeratin 19-positive hepatocellular carcinoma. Eur J Radiol 2019; 117:164\u0026ndash;170\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang X, Long L, Wei J, et al. Radiomics for diagnosis of dual-phenotype hepatocellular carcinoma using Gd-EOB-DTPA-enhanced MRI and patient prognosis. J Cancer Res Clin Oncol 2019; 145:2995\u0026ndash;3003\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang W, Gu D, Wei J, et al. A radiomics-based biomarker for cytokeratin 19 status of hepatocellular carcinoma with gadoxetic acid-enhanced MRI. Eur Radiol 2020; 30:3004\u0026ndash;3014\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Y, Chen J, Zhang Y, et al. Preoperative Prediction of Cytokeratin 19 Expression for Hepatocellular Carcinoma with Deep Learning Radiomics Based on Gadoxetic Acid-Enhanced Magnetic Resonance Imaging. J Hepatocell Carcinoma 2021; 8:795\u0026ndash;808\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu Q, Yu YX, Zhang T, et al. Preoperative Diagnosis of Dual-Phenotype Hepatocellular Carcinoma Using Enhanced MRI Radiomics Models. J Magn Reson Imaging 2023; 57:1185\u0026ndash;1196\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu L, Yang P, Yen EA, Wan Y, et al. A multi-organ cancer study of the classification performance using 2D and 3D image features in radiomics analysis. Phys Med Biol 2019; 64:215009\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCho ES, Choi JY. MRI features of hepatocellular carcinoma related to biologic behavior. \u003cem\u003eKorean J Radiol\u003c/em\u003e 2015 May-Jun; 16:449\u0026thinsp;\u0026ndash;\u0026thinsp;64\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoneda N, Matsui O, Kobayashi S, et al. Current status of imaging biomarkers predicting the biological nature of hepatocellular carcinoma. Jpn J Radiol 2019; 37:191\u0026ndash;208\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa WJ, Wang HY, Teng LS. Correlation analysis of preoperative serum alpha-fetoprotein (AFP) level and prognosis of hepatocellular carcinoma (HCC) after hepatectomy. World J Surg Oncol 2013; 11:212\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChan MY, She WH, Dai WC, et al. Prognostic value of preoperative alpha-fetoprotein (AFP) level in patients receiving curative hepatectomy- an analysis of 1,182 patients in Hong Kong. Transl Gastroenterol Hepatol 2019; 4:52\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang S, Xu L, Dai F, et al. Construction of a predictive nomogram and bioinformatic investigation of the potential mechanism of postoperative early recurrence of hepatocellular carcinoma meeting the Milan criteria. Ann Transl Med 2022; 10:866.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"dual-phenotype hepatocellular carcinoma, radiomics, CT, hepatocellular carcinoma","lastPublishedDoi":"10.21203/rs.3.rs-6285893/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6285893/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBACKGROUND.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDual-phenotype hepatocellular carcinoma (DPHCC), a highly aggressive subtype, is difficult to diagnose preoperatively based on morphological characteristics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOBJECTIVE.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe aimed to develop a radiomics model based on computed tomography (CT) data for non-invasive preoperative identification of DPHCC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMETHODS.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCT images and clinical data of 258 patients from Institution I were included in this study as a training set. Among them, 119 patients were diagnosed with DPHCC, and the rest 139 patients were treated as a control group. Radiomics features were extracted from regions of interest (ROI), and the features were selected by recursive feature elimination with cross-validation (RFECV). The logistic regression (LR) and random forest (RF) algorithm were used to develop four models to differentiate DPHCC and non-DPHCC, including the radiomics model, the clinical model, the radiologic model and the fused model. In addition, the effectiveness of the prediction models was evaluated by various indexes. CT images and clinical data of 58 patients from Institution II were included as an independent validation set. The models were evaluated on the independent validation set to assess the robustness of these models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRESULTS.\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAmong these models,\u003cstrong\u003e \u003c/strong\u003ethe radiomics model shows a balance of effectiveness and robustness. For the radiomics model, the mean area under the curve (AUC) of five-fold cross-validation on the training set reached 0.753, while the AUC on the independent validation set was 0.734. A radiomics signature containing 35 radiomics features was established for non-invasive prediction of dual-phenotype hepatocellular carcinoma. The fused model exhibits the best performance, with a training set AUC of 0.885 and an independent validation set AUC of 0.763.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCONCLUSION.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCT radiomics features of the tumor correlate significantly with the expression status of the bile duct phenotype in HCC. The radiomics feature, AFP level, ill-defined margin, pseudo-capsule and intra-tumoral necrosis are the essential features of the prediction model, potentially influencing the development of individualized treatment strategies for HCC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCLINICAL IMPACT.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe radiomics model based on computed tomography (CT) data contribute to non-invasive preoperative identification of DPHCC.\u003c/p\u003e","manuscriptTitle":"Non-invasive Radiomics Model for Preoperative Prediction of Dual-phenotype Hepatocellular Carcinoma Based on CT Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-02 10:39:52","doi":"10.21203/rs.3.rs-6285893/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4b895e4a-ddca-4d4d-a35c-32f9402b7a44","owner":[],"postedDate":"April 2nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-04-14T02:23:44+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-02 10:39:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6285893","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6285893","identity":"rs-6285893","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}