MRI Radiomics Combined with Clinicopathological Factors for Predicting 3-Year Overall Survival of Hepatocellular Carcinoma After Hepatectomy

Introduction

Liver cancer, the third most significant cause of cancer fatalities, has drastically affected lives and health globally.¹ Furthermore, the incidence of HCC is estimated to be between 85% and 90% of all primary liver cancers.² HCC is being diagnosed at an earlier stage with the advancement of medical technology. Surgical resection may be utilized in the HCC patients of early stage with the aim of achieving curative outcomes. However, it should be noted that only a relatively small percentage of patients undergoing this treatment will eventually experience adequate recovery. Precise prognosis prediction is crucial for developing effective treatment strategies. The risk of HCC cannot be classified or categorized based on the particular patient. Therefore, identifying high-risk patients following HCC requires a practical and consistent methodology.

Several clinicopathological features have been demonstrated to increase predicting the prognosis of HCC. Numerous diagnostic and prognostic models for HCC utilize alpha-fetoprotein (AFP), a widely regarded tumor marker associated with HCC.^3–8 The threshold for pathological alterations associated with HCC is generally defined as >20 ng/mL, and a concentration of >400ng/mL is generally accepted as one of the crucial diagnostic indicators for HCC.⁹ KI-67 is a marker of cellular proliferation and an independent predictor of HCC recurrence.^10–12 In addition, KI-67 can indicate tumor biology and determine prognosis.

Previous imaging studies^13–15 have relied on the qualitative assessment of tumor morphology, density and enhancement, which are laborious means of quantifying tumor information and susceptible to misdiagnosis due to the subjective influence of radiologists. Heterogeneous malignancy cells and their microenvironment have been shown to impact tumor biology, which can be examined by imaging. Radiomics, an emerging field, constructs a statistical model using multiple high-dimensional mineable variables derived from medical imaging data to facilitate diagnosis, prognosis and therapy monitoring.

The relationship between MRI-based radiomics, pathology and prognosis in HCC patients is poorly understood. The prognosis of liver cancer patients has not been assessed using MRI-based radiomics. Studies have focused on predicting recurrence-free survival in HCC patients,^{3,4,12,16–18} with fewer examining OS.¹⁹ Therefore, this research aims to construct a model using MRI-based radiomics and clinical metrics to predict 3-year OS in HCC patients. The purpose of the model is to facilitate clinical decision-making, treatment strategy and long-term postoperative follow-up. The primary objective of this research is to develop novel prospects for targeted precision therapy.

Materials and Methods

Patients

This retrospective study was approved by the Ethics Committee of Lishui Municipal Central Hospital (Approval No. 2023–720). As this was a retrospective study, informed consent was not required, nor were these decisions in any way detrimental to the rights of the study Patients. The confidentiality of patient data was strictly protected to ensure compliance with all relevant regulations and in line with the Declaration of Helsinki. According to the following inclusion criteria, 141 consecutive patients with liver cancer who underwent hepatectomy in the Department of Hepatobiliary and Pancreatic Surgery of Lishui Central Hospital, Zhejiang Province, China from September 2010 to July 2023 were enrolled: (1) patients who underwent primary radical hepatectomy and were pathologically confirmed to have HCC; (2) histopathological evidence showed no residual neoplastic tissue at the margin of the resection, which is required by R0 resection standards. (3) no lymph node and extrahepatic metastasis; (4) no major vascular invasion; (5) no radiofrequency ablation, transcatheter arterial chemoembolization (TACE) and chemotherapy prior to surgery; (6) patients who underwent enhanced MRI preoperatively. Exclusion criteria were as follows: (1) patients with benign or mixed liver tumors; (2) combined with other malignant tumors; (3) death due to surgical complications; (4) liver transplantation performed during the course of the disease; (5) incomplete clinical or follow-up data. Figure 1 illustrates the procedure for recruiting participants for the study. All patients underwent hepatectomy and pathological evaluation within two weeks of collection of clinical features and MRI examination. We randomly divided eligible patients into two groups in a 7:3 ratio: the training set (n =98) and the validation set (n =43).

Figure 1 Flow chart for screening patients.

Preoperative Clinical Characteristics

Data was collected from the electronic medical record system for the following patients. (1) The basic clinical characteristics of preoperative patients such as age, sex, body mass index (BMI), hypertension and diabetes mellitus; (2) preoperative routine laboratory indexes such as hepatitis B virus (HBV) status, alpha-fetoprotein (AFP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), γ-glutamyl transferase (GGT), direct bilirubin (DBIL), total bilirubin (TBIL), prealbumin (PALB), albumin (ALB), white blood cell count (WBC), neutrophil count (NEUT), lym monocyte count (MNC), SII (platelet × neutrophil/lymphocyte), ALBI (0.66 × log10 [bilirubin (μmol/L)] − 0.085 × [albumin (g/L)]); (2) Operation information such as intraoperative blood loss and blood transfusion; (3) pathological information such as tumor number, size, microvascular invasion (MVI), Ki-67; (4) tumor stage was determined based on the Child Pugh stage, the Barcelona Clinic Liver Cancer (BCLC)²⁰ classification system and the American Joint Committee on cancer’s tumor-node-metastasis (TNM)²¹ system. Histopathological analysis was performed by a pathologist with 10 years of experience in liver pathology, who was blinded to the radiological and clinical findings. MVI is defined as the microscopic observation of tumor cell nests within the wall of the portal vein, hepatic vein, and peritumoral vessel, lined by a single layer of endothelium.²² Immunohistochemistry staining for Ki-67 protein was performed. The Ki-67 expression was assessed by noting the percentage of positively stained cells with brownish yellow nuclei,²³ Supplementary Figure 1 shows the immunohistochemical results of MVI and Ki-67.

MRI Procedure

The German Siemens Aheral.5T or Siemens Symphony 1.5T superconducting MRI imaging system was used. MRI scans were performed using a T2-weighted sequence and a multiphase CE-MRI sequence. The T2-weighted sequence parameters were as follows: TR 3500 ms, TE 90 ms, FOV 380 mm × 380 mm, matrix 320 × 320; CE-MRI scans were performed with three-dimensional volume interpolation (3D-VIBE): TR 4.1ms, TE 1.8 ms, FOV: 380 mm × 380 mm, matrix: 320 × 320, thickness 5 mm, gap1 mm. The images of AP, PVP and DP were collected after injecting contrast agent Gd-DTPA (dose 0.1 mmol/kg, flow rate 2 mL/s) for 20–30 s, 60–70 s and 180 s, respectively.

Follow-Up

Patients with HBV were treated with anti-HBV according to protocols. All HCC patients were followed up at 3–6 months postoperatively for the first 2 years. The follow-up frequency was increased to 6–12 months. To detect recurrences of the tumor, liver and renal function, blood parameters and abdominal enhancement CT and chest CT were examined. Abdominal MRI or PET-CT were performed to detect recurrence. OS was assessed from the day of surgery until the time of death or the most recent follow-up.

Image Preprocessing and Tumor Segmentation

We exported T2WI and CE-MRI (AP, PVP and DP) scans in DICOM format from the picture archiving and communication system. MRI images were manually specified by tow radiologist blinded to the clinical information of the patients using Radcloud (http://radcloud.cn). Senior radiologists defined tumor boundaries if the variation was 5%.²⁴

Radiomics Feature Extraction

A total of 1688 quantitative imaging features were extracted from MR images. These features were grouped into three groups. Group 1 (first-order statistics) consisted of 126 features that quantitatively indicate the distribution of voxel intensities within the MRI image through commonly used and basic metrics. Group 2 (shape- and size-based features) contained 14 three-dimensional features that reflect the shape and size of the region. We classified 525 textural features into group 3 (texture features) using grey-level run-length and grey-level co-occurrence texture matrices.

Feature Selection and Model Construction

Various image features can be determined using the methods described before. A few features may not be applicable to each task. Therefore, task-specific characteristics selection and dimensionality reduction are important for effective performance. The first step was to select features for each volume of interest. The variance threshold, SelectKBest and least absolute shrinkage and selection operator reduced redundant functions. Next, BernoulliNB and MultinomialNB classifiers were employed to maximize radiomics algorithm discrimination for each sequence and combination to develop the most excellent radiomics model. To create the final fusion model, clinicopathological characteristics were included in the radiomics model and evaluated using ROC and Delong’s test. Nomograms were generated from fusion models. Clinical value was assessed using decision curve analysis, whereas nomogram prediction accuracy and differentiation were analyzed using calibration curves.

Statistical Analysis

Categorical variables were represented numerically and proportionately and contrasted using the x² or Fisher’s exact test. The x² test, Fisher’s exact test and logistic regression analysis were conducted using SPSS version 23.0, a statistical software developed by IBM Corporation. Furthermore, volume of interest segmentation, feature extraction and model construction processes were performed using the Radcloud platform (http://radcloud.cn/). Manual tumor segmentation, feature extraction, model creation and evaluation comprised the workflow (Figure 2). Finally, we plotted many datasets using R (4.3.1). Extracting and selecting features and performing statistical analysis are addressed in the Supplementary Tables 1–7.

Figure 2 Flowchart of radiomics analysis.

Results

Patient Characteristics

A total of 141 patients (male/female: 123/18; age <60/age>60: 86/55) were included in the final study group and randomly divided into the training set (n = 98) and validation set (n = 43), including the survival groups (n = 111) and non-survival groups (n = 30) based on 3-year survival after hepatectomy. Demographic and clinicopathological characteristics of included patients are shown in Table 1. Both training and validation sets showed no significant difference in characteristics (P = 0.087 to 0.999) including 3-year OS status.

Table 1 Baseline Characteristics of Training and Validation Groups

Correlation Between Clinical Factors and 3-Year OS in Patients

Univariate analysis revealed that five factors (AFP, Lym, largest tumor size, KI-67 and MVI) were significantly correlated with prognosis (P < 0.05). Multivariate analysis confirmed that KI-67 (p = 0.019, OR = 1.030, 95%CI: 1.005–1.056) and AFP (p = 0.002, OR = 2.887, 95%CI: 1.492–5.547) were independent risk factors for 3-year OS (Table 2).

Table 2 Clinical and Pathological Factors for 3-Year Survival in Patients with HCC

Predictive Performance of Radiomics Signatures from Different MRI Sequences and Classifiers

The Results of MRI sequences for variance threshold, SelectKBest, and LASSO are presented in Supplementary Tables 1 and the Table 3 shows the predictive performance of imaging radiomics features extracted from different MR sequences by various classifiers. We found the fusion radiomics model (T2+AP+PVP+DP) with the MultinomialNB classifier highly predictive in both the training and validation sets. The AUC for the training set was 0.899 (95% CI: 0.817–0.981) and for the validation set was 0.768 (95% CI: 0.615–0.921). We then incorporated independent clinical factors (KI-67 and AFP) into the fusion radiomics model (T2+AP+PVP+DP) to construct a predictive model using the variance threshold, SelectKBest and LASSO MultinomialNB. We first selected 2299 features from 6752 features using the variance threshold method, then 882 using select K best and 18 using LASSO. Figure 3 displays the model construction, while the other models can be found in the Supplementary Tables 2–7. The fusion model demonstrated superior predictive accuracy with an AUC of 0.910 (95% CI: 0.823–0.997) in the training dataset and 0.846 (95% CI: 0.687-1.000) in the validation set. The fusion model outperformed the other model in identifying 3-year OS status in the training set (T2 or PVP or DP; MultinomialNB or BernoulliNB classifier) or validation set (T2+AP+PVP+DP by MultinomialNB or BernoulliNB classifier; T2+AP+PVP+DP+AFP+Ki-67 by BernoulliNB classifier) (DeLong’s test, P < 0.05). Although there was no statistically significant difference in identifying 3-year OS status between the training set (fusion model) and the training set (AP; MultinomialNB classifier) (DeLong’s test, P =0.066) or (BernoulliNB classifier) (DeLong’s test, P = 0.062), it was close to thresholds. Given all the above, we finally selected the fusion model as our best model. Furthermore, the fusion model outperformed BCLC, TNM and Child Pugh models in 3-year OS prediction (DeLong’s test, P < 0.05) (Figure 4).

Table 3 The Predictive Performance of Imaging Radiomics Features E×tracted from MR Sequences by Various Classifiers

Figure 3 The construction process of the fusion model. (a) Variance threshold on feature select. (b) Select K best on feature select. (c) Lasso algorithm on feature select. (d) Lasso path-; MSE path. coefficients in Lass model. Coefficients in the Lasso model. (e) Using the Lasso model, 18 features which are correspond to the optimal alpha value were selected. (f) Heatmap depicting the correlation coefficient matrix in the radiomics model.

Figure 4 The ROC curve of different models predicts 3-year OS. Abbreviations: TNM, the American Joint Committee on Cancer tumor-node-metastasis system; BCLC, Barcelona Clinic Liver Cancer.

3-Year OS Predictive Nomogram Construction and Evaluation

Clinicopathological and radiomics predictions were used to develop a 3-year OS model using AFP, Ki-67 and R-score risk factors. More importantly, Figure 5 shows the visual presentation of this model with R-score, AFP level and Ki-67. The calibration curves indicate that the predicted and actual 3-year OS in the training set are similar (Figure 6). The nomogram provided superior clinical performance for predicting 3-year OS in the training and validation set (Figure 7).

Figure 5 Nomograms for predicting the 3-year OS of HCC patients.

Figure 6 Calibration curves for training and validation sets.

Figure 7 Decision curves for training and validation sets.

Discussion

Based on MRI radiomics and clinicopathological information, a predictive model was developed for HCC patients of 3-year OS status after hepatectomy. Based on this model, we derived excellent and stable biases of 3-year OS in HCC patients, which is essential for timely postoperative intervention and improves long-term survival. Furthermore, it can be used for the rapid delivery of postoperative interventions and screening suitable participants for clinical trials involving postoperative treatments is crucial in enhancing long-term survival. To the best of our knowledge, predictive models based on MRI radiomics scores and clinicopathology data have not been used to estimate the 3-year OS of HCC patients who underwent hepatic resection.

MRI is more effective than CT for detecting and predicting the prognosis of early-stage HCC, which can appear at multiple parameters and sequences.²⁵ The method has evolved into primary imaging and standard preoperative examinations for HCC.²⁶ Several MRI-based radiomics models have been developed to enhance the diagnosis and prognosis of patients with HCC. Kim et al³ developed a clinicopathological-like recurrence prediction model using CE MRI data from 167 patients, exhibiting clinical significance. Furthermore, a radiomics nomogram developed by Zhang et al⁴ to predict HCC recurrence combined with traditional clinical radiology risk factors performed better than conventional nomograms and demonstrated the importance of MRI radiomics. A meta-analysis²⁷ of 15 studies involving 981 patients indicates that MRI radiomics holds high accuracy in predicting MVI of HCC and can be considered a non-invasive method for evaluating the presence of MVI in HCC patients. Zhang et al²⁸ assessed MRI scans of 136 primary liver cancer patients and found that radiomics characteristics were independently associated with OS of distinct pathological types. Different MRI sequences contain different information and possess varying degrees of predictive efficacy for progression of HCC. Wang et al¹⁹ established an MRI-based radiomics model to predict the 5-year OS of HCC patients. It was found that the dynamic contrast-enhanced imaging sequence showed the highest level of predictive accuracy.

An extensive effort was made by comparing different sequences and classifiers, our study found that the fusion model fusing T2, AP, PVP, DP sequences and independent clinicopathological factors of AFP and Ki-67 showed superior predictive capability. T2WI shows tumor spatial dimension and AP, PVP and DP sequences indicate tumor blood vessel and cell activity. Therefore, fusion radiomic models utilizing T2, AP, PVP and DP sequences may precisely exhibit tumor characteristics, improving diagnostic and therapeutic outcomes. The multivariate analysis showed that KI-67 and AFP were independent risk factors for 3-year OS (Table 2). The 18 radiomatic features of the fusion radiomic model (4 T2 features; 6 AP features; 4 PVP features; 4 DP features) implied tumor biological behaviour and tumor microenvironment. The reason why the AP sequences extracted the most features may be that the hepatic artery serves as the primary source of HCC supply and the AP sequence could more clearly reflect microvasculature alterations of HCC. Furthermore, HCC is a solid tumor rich in blood vessels and angiogenesis plays a significant role in its prognosis.²⁹ Arizumi et al³⁰ demonstrated that decreased arterial enhancement during sorafenib treatment was associated with the OS of HCC patients. Zhu et al³¹ found that PVP and DP were less predictive than AP. GLZM_ZoneEntropy is the zone entropy in the gray level size zone matrix, which is a texture feature. Higher values indicate more significant texture heterogeneity. Intriguingly, we found that GLSZM_ZoneEntropy in T2, AP and DP sequences are all positively associated with 3-year OS in HCC patients. Accordingly, our findings are in line with the widely accepted cognitive belief that an increase in texture heterogeneity signifies a poor prognosis.³²

In this study, AFP and KI-67 were independent clinical and pathological risk factors of OS. Consistent with previous research, these clinicopathologic factors have been established as effective predictors of clinical outcomes.^6,10–13,28 The study results indicate no notable correlation between MVI, tumor number, tumor size and survival outcomes. In this study, the relatively short follow-up period and the limited number of cases could explain this observation.

Advantages

This was the first study to combine clinical, pathological and MRI texture features to predict 3-year postoperative OS in HCC patients. Their predictive performance was improved than the BCLC stage, TNM stage and Child-Pugh class. Second, different machine learning methods were utilized to compare training and validation set to improve model accuracy and objectivity. This study mainly established techniques and methodologies for computer-aided and artificial intelligence analysis to predict 3-year postoperative OS in patients with HCC. We developed a nomogram to predict the 3-year OS of HCC patients using clinicopathological and radiomics factors.

Limitations

There are several limitations in our research. Specifically, selection bias was unpredictable in a single-center retrospective analysis with a small sample size. We analyzed MRI sequences using different classifiers for more detail, but we needed external validation with larger datasets from other facilities. Manual segmentation was employed for selection, which is more precise than semiautomatic and automatic segmentation. However, unclear image borders or subjective errors can result in segmentation errors, which makes subjectivity invariable. Furthermore, retreatment following recurrence and other factors can have a negative impact on 3-year OS prediction.

Conclusion

In summary, our predictive model for the 3-year OS of HCC after hepatectomy using MRI radiomics signatures and clinicopathological variables showed promising findings. Therefore, our findings reveal that this nomogram model may improve clinical decision-making and post-surgery surveillance.