A Cox Nomogram for Assessing Recurrence Free Survival in Hepatocellular Carcinoma Following Surgical Resection Using Dynamic Contrast‐Enhanced MRI Radiomics

The prognosis of hepatocellular carcinoma (HCC) is difficult to predict and carries high mortality. This study utilized radiomic techniques with clinical examinations to assess recurrence in HCC.

H epatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality and is the sixth most common in incidence worldwide. 1Despite partial hepatectomy being one of the most effective treatments for HCC, postoperative recurrence is a major risk to long-term survival. 2,3The basis of intrahepatic recurrence is likely due to intrahepatic metastasis of the primary tumor, or a de novo emergence of a new tumor. 4,5Recurrences with the latter mechanism generally have a poorer prognosis, and a higher rate of recurrence within 2 years. 6It is therefore important to identify factors that may precede early recurrence to allow for appropriate patient triaging and treatment planning.
Highly invasive differentiation, microvascular invasion (MVI), and satellite lesions have been identified as pathological factors associated with early recurrence of HCC. 3,7,8A history of cirrhosis or fibrosis is often a key factor causing remnant liver multicentric recurrences. 9Pathologically, the peritumoral parenchyma is representative of cancerous heterogeneity. 61][12] MaVI frequently seen in the portal vein is detrimental to prognosis. 13Detection of MaVI's using dynamic contrast-enhanced MRI (DCE-MRI) has been highly beneficial.
Radiomic approaches offer a high level of information regarding tumor characteristics by converting traditional medical images into high-dimensional, quantitative and mineable data sets. 135][16] However, most studies have only investigated finite peritumoral or intratumoral regions [17][18][19] and have not been extended to cover the volume of interest (VOI) of the tumor and its periphery (within 10 mm) and the background liver parenchyma.
Thus, the aim of this study was to combine tumoral, peritumoral, and background liver regions in DCE MR images of HCC without MaVI to develop a radiomic model for preoperatively assessing recurrence-free survival (RFS).To add clarity, three unique models were created and evaluated: 1) the total radiomic score, which comprised data from all 3 VOI's, 2) the clinical-radiological model, which incorporates elements of clinical judgment with radiologic imaging, and 3) the combined radiomic model, which comprised the COX nomogram, total radiomic score, clinical judgment, and MRI imaging.

Patients
This retrospective study was approved by our institutional ethics review board, and the requirement for informed consent was waived.
This retrospective study was carried out in three centers: center A, center B, and center C. In total, 249 patients were enrolled.After combining data, enrolled patients in center A and center B were assigned to the training cohort and the test cohort.The patients in center C were assigned to the external validation cohort.
There were 249 HCCs undergoing surgical resection at the three institutions between January 2016 and October 2020.Inclusion criteria were 1) patients with pathologically proven HCC; 2) patients who underwent preoperative abdominal contrast-enhanced MRI examination within 1 month before surgery; and 3) surgical removal of a single or multiple solitary HCC lesions in one liver lobe according to the Guidelines for the Diagnosis and Treatment of Hepatocellular Carcinoma (2019 Edition).Exclusion criteria included: 1) patients who had not undergone treatment (eg transarterial chemoembolization [TACE], ablation therapies, etc.) before surgery; 2) patients with MaVI-induced gross bile duct tumor thrombosis (MaVI refers to cancerous emboli in the gross vein and its main branches, which are often detectable by preoperative imaging) 19 ; 3) insufficient clinical or pathological information; and 4) images in which key experimental parameters could not be determined/utilized due to image quality.Figure 1 shows a flow diagram of patients included and excluded from the study.

Follow-Up
In the first month after surgery and every 3 months thereafter, contrast-enhanced CT and MR imaging studies were performed.Early recurrence was defined as intrahepatic and/or extrahepatic recurrence within a 2-year period following HCC tumor resection.After 2 years, patients underwent follow-up every 6 months.End points examined were RFS, which was defined as the time from the date of surgery to the early recurrence.The last census point was October of 2022 for living patients.

Clinical-Radiological Variables
An independent review of the radiological features was performed by three abdominal radiologists (A = YHR, B = ZQ, and C = CXS) with 5, 10, and 20 years in diagnostic radiology experience, respectively, who were blinded to clinical data and imaging reports but were aware that the patients had HCC.Disputes between readers were discussed until an agreement was reached.In patients with multifocal lesions, the largest lesion was assessed.Evaluations were conducted on the following imaging features: 1) number of tumors detected (solitary vs. multiple); 2) the size of the main lesion 20 (≤5 cm vs. >5 cm); 3) margin features 11 (smooth vs. irregular); 4) enhancing capsule sign 21 (absent vs. present); 5) peritumoral arterial phase-contrast enhancement 22

Segmentation of Tumor Images
The DCE-MRI (AP, VP, and DP) images were exported as DICOM format and loaded into the open-source software 3D slicer (version 4.10.2;www.slicer.org)for manual segmentation.Using the image with the largest cross-sectional tumor area and two adjacent images, radiologist A manually segmented the tumor VOI (VOI tumorInner ).The tumor border plus 1 cm in each direction along with the internal tumor area was also reconstructed by topology and defined as VOI Plus .A background VOI (VOI Background ) was manually defined as a region outside of the segmented VOI in the liver parenchyma.To optimize segmentation workload, only the AP images were delineated manually in our study, and the DP and VP images were registered by General Registration (Elastix).This software is a fast deformable registration toolbox for 3D medical images based on the open source software of image segmentation and registration toolkit ITK (https://elastix.lumc.nl/). 24Finally, manual corrections were made to the image registrations and contours on the VP and DP images.An overview of the tumor segmentation workflow was illustrated in Fig. 2.

Radiomic Feature Extraction
MRI signal intensity values vary based on acquisition parameters, which impact the extracted radiomic features.Image normalization was carried out using a method that remaps the histogram to compensate for variations arising from scanner manufacturer and magnetic field strength.The voxels within the tumor region with intensities outside the range of μ AE 3σ (μ and σ represent the mean and standard deviation of the image intensities within the tumor region, respectively) were rejected.Radiomic extraction was performed using Pyradiomics (version: 2.12; https://pyradiomics. readthedocs.io/en/2.1.2/).The image was resampled to a pixel pitch of 3 Â 3 Â 3 mm using the B-splines method to counteract the interference caused by uneven spatial resolution.Afterward, the images were preprocessed with wavelet filters or Laplacian filters based on a variety of parameters.These included first-order statistics, shape, and textures features.The radiomic features were normalized with the Z-score.Radiologist A (Y.H.R., with 5 years of experience) manually segmented all patients VOIs.A random selection of 30 cases were selected for repeated feature extraction by radiologist A and by radiologist B (Z.Q., with 10 years of experience) to test feature stability.The intraclass correlation coefficient (ICC) was used to evaluate the interobserver reproducibility of features.Radiomic features with ICC values greater than 0.75 (indicating good agreement between observers) were selected for further analysis.

Statistical Analysis
To select the radiomic features with high value prediction ability from the ICC-determined radiomic features with high stability, the least absolute shrinkage and selection operator (LASSO) Cox regression analysis was performed. 24In order to reduce overfitting and align the parameters, 10-fold cross-validation was performed.Radiomic scores were calculated for each patient using a linear combination of radiomic features weighted by their LASSO coefficients.
A univariable and multivariable Cox regression analysis was conducted on the training cohort to evaluate independent predictors of RFS.Multivariable Cox regression analysis was conducted with variables with P values less than 0.05 in the univariable analysis.The independent risk factors of all models were gradually selected according to the Akaike information criterion to build the best model. 25The proposed model's discriminative ability was measured using Harrell's concordance index (C-index) in the training cohort and confirmed in the internal and external validation cohorts.Observed RFS was compared with model-predicted probability using calibration curves.Based on the total radiomic score, and the clinical-radiological model, a dataset was constructed.The clinicalradiological model was constructed using the four factors (AFP, Size, Satellite nodules, and cirrhosis), which were then also used in the combined radiomic model.
An ensemble learning model was built based on Kmeans clustering. 26Cluster analysis categorized two risk levels (high and low)

Results
Methodology A total of 249 patients were identified for the study and were divided into the training (n = 153), test (n = 66), and external validation (n = 30) cohorts.Early recurrence was found in 119 (47.8%) of patients with 73, 31, and 15 in the training, test, and external validation cohorts respectively.Across all three cohorts, there was no significant difference in recurrence rate (P = 0.994).The clinical parameters for each patient are summarized in Table 2.No significant difference was detected for any characteristic between the training, test, and external cohorts (P = 0.072-0.994).Among the radiomic features extracted from DCE-MRI AP, VP and DP images, only features with inter-observer correlation coefficients higher than 0.75 were considered.The LASSO feature selection method was then used to optimize model parameters.In the training cohort, we selected non-zero coefficient features from multiple phases, MRI images, and implemented them quantitatively into three radiomic scores based on VOI tumorInner , VOI plus , and VOI background .As a final step, all significant radiomic features were combined into a single combined radiomic score.The coefficients of each feature in the radiomic score and formulas are shown in the Supplementary File S1.

Construction of the Combined Radiomic Model
According to the univariable analysis, six factors significantly impacted RFS, including two clinical variables (AFP, Cirrhosis) and four radiology features (irregular margin, peritumoral arterial phase contrast enhancement, tumor size and satellite).Multivariable Cox regression analyses determined AFP, cirrhosis, tumor size, and satellite as independent predictors of RFS (Table 3).These factors were selected to build the clinical-radiological model.From the coefficients in the multivariable COX regression, the clinical-radiological model and the three radiomic scores based on the three VOIs were combined to build the RFS COX nomogram (Fig. 3).The calibration curves for the nomogram in assessing 24 months survival rate after surgery in the training, test, and external cohorts are shown in Fig. 4a-c, respectively.

Risk Stratification by kmeans Cluster Analysis
Patients were classified as low-risk or high-risk patients based on kmeans cluster analysis defined from the three radiomic scores and combined radiomic model score as cluster centers.6 illustrates that the training cohort showed a significant correlation between high-risk patients and shorter postoperative survival (Fig. 6a), which was also confirmed in the test cohort (Fig. 6b) and the external cohort (Fig. 6c).

Discussion
In the current study, a radiomic model was built from dynamic contrast-enhanced MRI radiomic features.According to the recurrence and nonrecurrence status and survival data of patients within 2 years, the model was trained to screen the most predictive imaging features.According to the test set and external verification set, the model was verified to assess the RFS of patients with hepatocellular carcinoma after radical resection.In the current study, radiomic signatures in contrast-enhanced MRI images were used as risk markers of tumor recurrence.The model was trained according to the patients' recurrence status and survival data within 2 years.The radiomic features with the greatest predictive value were screened, and the model was verified according to the test set and external validation set to predict the survival rate of patients with hepatocellular carcinoma after radical resection. 27here are several aspects of tumor biology that could be reflected by these features, including the underlying textural information. 28An intuitive understanding of the correlation between one radiomic feature and biological behavior is difficult. 29,30We further used all features from three radiomic scores (one for tumor VOI, one for tumor and periphery ROI, and one for background ROI) and combined all radiomic scores into the total radiomic score.The correlation between MVI radiological features of the peritumoral region (including nonsmooth tumor margins, peritumoral enhancement and satellite nodules) has been demonstrated by recent studies. 6,23here is a high rate of invasion of tumor cells within the peritumoral parenchyma. 11To stratify patient risk and improve long-term survival, it is important to identify predisposing predictors for early recurrence before surgery. 29,30To this end, we first created a clinical-radiological model using peritumoral hallmarks and behaviors to assess early recurrence.The results demonstrated that satellite instability, cirrhosis, AFP, and tumor sizes >5 cm independently impaired RFS.However, for the factors of smooth margin appearance and peritumoral enhancement, our univariable analysis showed significant associations with RFS, but the multivariable analysis did not confirm these results.This differs from prior reports. 10,19,31Considering that our radiomic analysis used strict inclusion criteria, inconsistencies may be related to selection bias.Further, the satellite nodules had a higher predictive value than other radiologic features in predicting RFS.There has been some evidence that satellite nodules may reflect the heterogeneity of the peritumoral parenchyma in previous studies. 8,32MVI firstly involves the peritumoral parenchyma, and functions in the hematogenous dissemination pathway of peritumoral enhancement, satellite nodules, MaVI, intra-and extrahepatic metastasis. 11,12Satellite nodules were highly associated with heterogenous angiogenesis and HCC than peritumoral enhancement and were independent of DCE-MRI as determined herein.This may explain why peritumoral enhancement was not a significant factor in assessing RFS.In our study, AFP, the maximum tumor diameter, and cirrhosis were independent predictors for RFS in the clinical-radiological model.Chan et al [21][22][23][24][25][26][28][29][30][31][32][33][34][35][36] showed that early recurrence was significantly associated with a high AFP and tumor sizes greater than 5 cm. As repted by Kim et al, 37 early recurrence was significantly linked to cirrhosis, which was consistent with our results.Further, tumors larger than 5 cm were associated with an increased incidence of satellite nodules, elevated AFP levels, fibrosis, and cirrhosis.38 Since most HCCs develop from hepatic inflammation secondary to pathologies, the prognostic predictor should consider both tumor characteristics and impairments in hepatic function such as cirrhosis.The combined radiomic model showed similar performance to the total radiomic score and the clinical-radiological model, while the relative weights of each were different, indicating that the contribution of the total radiomic score and clinical-radiological model were independent and essential to the model.
Utilizing computational-based radiologic imaging, radiomics has shown great potential in assessing the course of HCC in previous studies. 39With radiomics, we were able to capture tumor and nontumor details by manually outlining the corresponding ROIs, to assess the accuracy between these ROIs.Models for RFS prediction could be efficiently built by selecting and integrating these features.Most radiomic studies have mainly focused on the internal and peritumoral expansion region of the tumor. 10,19,40To determine radiomic features, we proposed a new approach.
In the creation of the total radiomic score, three distinct regions were measured.The intratumoral region (VOI tumorInner ) region initiates from the core of the tumor to the outer perimeter of the tumor.The (VOI plus ) region initiates from the tumor perimeter to 1 cm beyond the tumor perimeter in all directions.Finally, the (VOI background ) represents the background liver signal not encompassed by either of the other two sections.In combination, these three regions comprised the total radiomic score.In the univariable analysis, the three corresponding radiomic scores were significantly associated with RFS and showed comparable prognostic performance in training and test cohorts.The radiomic parameters from the background VOI showed high prediction accuracy in the training and test cohorts which suggests that background parameters are indicative of patient survival and may provide important prognostic data for metastatic potential.
This study analyzed the ROIs of the largest crosssectional tumor area image and its two adjacent images.As such, it provided a more comprehensive evaluation of the lesion than previous 2D analysis. 15In addition, the peritumoral expansion VOI was more informative than liver background and intratumoral VOIs.This finding may be because the algorithm combined high-definition images of the peritumoral domains and spreading scopes of MVI.As such, the finding is consistent with previous investigation. 10ditionally, the total radiomic score from the three VOIs was superior to the clinical-radiological model for RFS.After combining the radiomic features into the combined radiomic model, an improved performance was achieved in training, test, and external cohorts.Accordingly, multiscale radiomic features involving a variety of regions have an advantage over single-scale and single-region radiomic features when assessing RFS in patients.The combined radionic model also had adequate calibration and discrimination.This model had a higher C-index than the other prediction models, which indicated a strong predictive value for RFS in HCC patients without MaVI following resection.
For clinical treatment, stratifying individuals according to risk is essential to triaging.We have proposed stratifying patients' risk level of RFS based on the Kmean.This method is more precise and accurate than relying solely on the cut-off median value. 37Using this approach, high-risk survivors could be differentiated from low-risk survivors using the proposed risk level cluster.It is proposed that the patients with high-risk recurrence should undergo more aggressive surgical and clinical therapy.
The present study was a multicenter study.The performance of the combined radiomic model was robust in the test validation cohort (from the same two centers [A and B] as the training cohort) and slightly lower in the external validation cohort (from center C).Although MRI data acquired at different field strengths were used in our study, the Cox nomogram results were not significantly affected.We suggest that this may be the result of several factors: first, grayscale standardization and resampling of images were applied to MR data before ROI segmentation; and second, we merged the two main centers' data of 3.0 T and 1.5 T data to one dataset cohort.In addition, as patients are often imaged in multiple MR scanners, with different vendors, field strength and imaging parameters, this may reflect clinical reality.As demonstrated in the external validation cohort, this approach has the potential to improve the assessment of RFS in HCC patients.

Limitations
Firstly, 200 cases are a small sample size for survival the analysis of radiomic nomograms.Secondly, using multicentric DCE-MRI datasets for space related features extraction can predispose results to differences in scanning slice thickness and acquisition parameters.Additionally, we extracted features from the image showing the largest tumor crosssectional area and two adjacent images, rather than applying a true volumetric approach.This could result in incomplete information.
Another major limitation regards the fact that images obtained for this study were from different field strengths and different machines.However, several steps were taken to mitigate the impact of this discrepancy.Firstly, all the original images collected in this study have been processed by the Image Biomarker Standardization Protocol (IBSI).Before ROI segmentation, the original magnetic resonance images were normalized by remapping histogram, and the average and standard deviation σ of voxel gray allocated to ROI were calculated.Subsequently, voxels beyond the range [3σ, +3σ] are excluded from the gray range.This reduces the influence of abnormal gray values on the model.Second, due to the heterogeneity of tumors, texture feature sets generally require interpolation of isotropic voxel spacing.However, voxel interpolation can affect image feature values as many image features are sensitive to changes in voxel size.Therefore, it is important to maintain consistent isotropic voxel spacing between different measurements and devices for reproducibility.We used the Bsplines method to sample 3.0 mm Â 3.0 mm Â 3.0 mm on the image for image resampling and interpolation processing.The slice thicknesses of all images were above 3 mm.After preliminary verification, the model was downsampled to 1.0 mm Â 1.0 mm Â 1.0 mm for feature extraction.While this study combined the data of 3.0 T and 1.5 T centers into one data set for analysis, the reality of clinical practice is that images are often obtained from the first available scanner regardless of imaging parameters.

Conclusion
This study shows that radiomics is a potential tool for the development of COX nomograms which have potential to be used for guiding the treatment of critical illnesses such as HCC.Specifically, the combined radiomic model proposed herein, which combines the total radiomic score with the clinical-radiological model showed great potential in assessing the incidence of RFS in HCC patients.This study is a pilot study of a large multicenter study, and future studies are planned to examine the relationship between HCC, recurrence, and radiomic features in greater detail.

FIGURE 2 :
FIGURE 2: Diagram shows the workflow of the tumor segmentations. 1) Image acquisition.First step: The DICOM format of contrastenhanced MR images in arterial, venous, and delayed phases (AP, VP, and DP, respectively).2) Delineated on AP.Second step: radiologists manually delineated the tumor on the AP image with the largest tumor cross-section and the two adjacent images to generate VOI tumorInner .3) Image segmentation.Third step: The border of the tumor plus 1 cm in each direction, along with the inner portion of the tumor was defined as the VOI Plus .A region outside the VOI was defined as the VOI background .The VOI's were processed for manual correction.4) Register on VP and DP.Fourth step: The AP VOIs were registered to both DP and VP images.

FIGURE 3 :FIGURE 4 :
FIGURE 3: The Cox nomogram for the combined radiomic model for recurrence-free survival was developed.With the training cohort as a reference, a nomogram was created, incorporating the 3 VOI radscores and the clinical-radiological model, scaling the model with the proportional regression coefficients.

Figure 5
Figure5shows high and low recurrence risk levels as different color dots (overall patients from training cohort, test cohort, external cohort).Figure6illustrates that the training cohort showed a significant correlation between high-risk patients and shorter postoperative survival (Fig.6a), which was also confirmed in the test cohort (Fig.6b) and the external cohort (Fig.6c).

Figure
Figure5shows high and low recurrence risk levels as different color dots (overall patients from training cohort, test cohort, external cohort).Figure6illustrates that the training cohort showed a significant correlation between high-risk patients and shorter postoperative survival (Fig.6a), which was also confirmed in the test cohort (Fig.6b) and the external cohort (Fig.6c).

FIGURE 6 :
FIGURE 6: Based on the Kaplan-Meier survival analysis of patients in the training cohort (a), test cohort (b), and external cohort (c), the recurrence-free survival of subgroups classified by kmeans clustering method.

TABLE 2 .
Training, Test, and External Cohort Characteristics the training dataset.The Kaplan-Meier method was then used to determine the risk levels in the training, test, and external validation cohorts.An analysis of two-sided log-rank tests was conducted to compare survival curves created with the Kaplan-Meier method.All statistical analysis was performed using R software (version 3.5.2R Foundation for Statistical Computing, Vienna, Austria).A two-sided P < 0.05 was considered statistically significant throughout the study. in

TABLE 4 .
Performance of the Survival Models