This article is mentioned in:

Introduction

The National Cancer Center report on cancer incidence and mortality in China in 2022 indicated that colorectal cancer (CRC) is the second most prevalent tumor type, following lung cancer (1). Notably, the liver is the most common site of metastasis for CRC and liver metastasis is a leading cause of mortality in patients (2). The management of liver metastases in CRC poses a key challenge in current treatment practices. Numerous patients with CRC have been reported to possess micrometastases in surgically resected liver specimens, which are considered invisible to the naked eye and imaging examinations before surgery, and these micrometastases often develop into metachronous liver metastases (MLM) (3–5). Consequently, MLM is prone to being overlooked during follow-up, which leads to treatment failure (6).

Compared with other treatments, such as radiotherapy and chemotherapy, tumor ablation notably improves patient prognosis and survival (7); however, current imaging methods such as CT, MRI and ultrasound, can occasionally miss microscopic lesions, which delay diagnosis and treatment. Consequently, the establishment of robust screening protocols is necessary to identify patients with CRC at an elevated risk of developing MLM, thereby facilitating early detection, which would enable timely surgical intervention, markedly improve prognosis and enhance overall survival in patients (8). Previous studies have explored various clinical factors, such as carcinoembryonic antigen (CEA) levels, lymph node status, KRAS and primary tumor site, as potential biomarkers for MLM; however, a clear consensus regarding these biomarkers has not yet been established (9,10).

Radiomics is a novel technique that extracts quantitative features from medical images and transforms the features into mineable data. Through radiomics, key imaging biomarkers that are undetectable to the naked eye can be identified (11,12). A key advantage of radiomics is the ability to analyze medical images across different modalities (13). The proposed methodology synergistically integrates multimodal data acquired from diverse imaging modalities, which effectively overcome the constraints inherent in single-modal analysis while facilitating a robust evaluation of potential synergistic benefits. By adopting a cross-modality approach, radiomics maximizes the use of extracted imaging information and provides a more comprehensive understanding of the underlying pathology (14,15).

We hypothesized that a change in the liver microenvironment may occur before the appearance of imaging manifestations. Previous studies that have investigated radiomics in patients with CRC predominantly focused on a unimodal approach or on multiple sequences with a single modality (8,16,17). To the best of our knowledge, there are a limited number of literature reports on the application of multimodality in patients with CRC. Therefore, the present study aimed to develop an integrated model combining CT and MRI modalities to improve the prediction of MLM in patients with CRC. The primary objective of the present study was to establish an imaging-based foundation for clinical prediction and prevention of MLM in patients with CRC, which may potentially prolong patient survival in the future.

Materials and methods

Subjects

The present study retrospectively analyzed a total of 1,144 patients with CRC clinically diagnosed at Sir Run Run Shaw Hospital (Zhejiang, China) between January 2010 and December 2020. The present study received approval from the Ethical Review Committee of Sir Run Run Shaw Hospital, Zhejiang University School of Medicine (approval no. 0465-2022; Zhejiang, China) and a waiver of consent was obtained due to the retrospective nature of the present study. A total of 157 cases were included in the present study, which included 67 patients with liver metastases within 2 years of treatment and 90 patients without liver metastases at postoperative follow-up. Clinical data of patients in terms of age, sex, carcinoembryonic antigen (CEA), carbohydrate antigen 199 (CA199), anal distance, diameter, tumor stage, circumferential resection margin (CRM) and extramural vascular invasion (EMVI), were collected from both MLM and non-MLM groups. Tumor staging followed the 8th American Joint Committee on Cancer Tumor-Node-Metastasis classification, whereas EMVI and CRM were assessed according to the 2017 European Society for Medical Oncology guidelines (18,19). Tumor dimensions and location were obtained from preoperative colonoscopy reports. All data analyzed were the characteristics available in the medical records of patients and were verified by an experienced radiologist (>10 years of experience).

The inclusion criteria for the present study were as follows: i) Confirmation of CRC diagnosis through biopsy or surgical pathology, ii) absence of primary tumors originating from other sites, iii) absence of evident signs of liver metastasis upon abdominal CT or MRI enhancement at the time of diagnosis and iv) development of liver metastasis within 2 years of treatment, confirmed through pathology or imaging review. The exclusion criteria consisted of: i) Postoperative patients from other hospitals; ii) combination of primary tumors from other sites; iii) presence of liver metastasis at the time of diagnosis; iv) lack of a 2-year follow-up period; and v) blurred image quality with sequences missing, in which the tumor could not be identified (Fig. 1).

Figure 1. Flow chart of the present study cohort.

Equipment

MRI acquisitions were performed using the following 3.0-T MRI scanners: Discovery™ MR 750 (GE Healthcare), Signa™ HDx (GE Healthcare) and MAGNETOM Skyra (Siemens Healthineers). The images of High-Resolution T2-weighted imaging (HRT2) were specifically utilized for lesion outlining. No intravenous contrast agents were administered.

The following CT scanning equipment were used: LightSpeed VCT (GE Healthcare), Optima CT620 (GE Healthcare), Revolution™ Maxima (GE Healthcare), Definition AS (Siemens Healthineers), Definition AS 40 (Siemens Healthineers), FORCE CT (Siemens Healthineers), Sensation16 (Siemens Healthineers), SOMATOM Definition Flash (Siemens Healthineers), SOMATOM go.Top® (Siemens Healthineers), UIHCT (United Imaging) and Definition AS+ (Siemens Healthineers). Intravenous contrast iohexol were administered.

Image segmentation and feature extraction

The liver was automatically outlined using the Livermask package (version 1.4.1; http://github.com/andreped/livermask). ITK-SNAP (version 3.8.0; http://www.itksnap.org) was then used to modify and annotate the bowel MR images and the automatically outlined liver CT images. All segmentation masks were reviewed by a junior radiologist (with >5 years of experience in radiology) and finally confirmed by another senior radiologist (with >10 years of experience in radiology). Disagreements were resolved by consensus decision making.

The International Biomarker Standardization Initiative standard (https://theibsi.github.io/) was used during the experimental procedure, z-score normalization was performed on the signal intensity of HRT2 images and histogram normalization was performed on CT images prior to feature extraction. The PyRadiomics package (version 2.1.2; http://github.com/AIM-Harvard/pyradiomics) was employed for radiomics feature extraction. To refine the feature set, features with low variance (P<0.01) were eliminated by conducting one-way ANOVA. An independent-sample t-test was conducted to estimate the radiomics features. Features with P<0.05 were considered significant for model development. The image acquisition and segmentation are shown in Fig. 2, and the radiomics feature extraction is shown in Fig. 3.

Figure 2. Image acquisition and segmentation. The ROI of the tumor was segmented and reconstructed in (A) and. (C) The ROI of the liver was segmented and reconstructed in (B) and. (D) ROI, region of interest.

Figure 3. Radiomics feature extraction. GLSZM, gray level size zone matrix; GLCM, gray level co-occurrence matrix; NGTDM, neighboring gray tone difference matrix.

Feature selection and model development

Since the imbalance in the original dataset (n=67 with MLM and n=90 without MLM) may lead to overfitting, a naive random over-sampling method was employed within the open-source Python package Imbalanced-learn (version 0.9.0, http://imbalanced-learn.org/) to create a balanced dataset (n=90 with MLM and n=90 without MLM). Missing information regarding the clinical characteristics (e.g., CEA and CA199) were replaced with the mean value of the corresponding features. Then bootstrap resampling was performed to derive confidence intervals and improve the stability of the model. With a balanced dataset, the Least Absolute Shrinkage and Selection Operator (LASSO) was utilized to select the optimized subset of features from the pre-processed features. For LASSO regression, α parameters of 0.093871 for CT, 0.061385 for MRI and 0.030567 for clinical features were used. A prediction model was then built using the Random Forest (RF) algorithm. The RF was built with 500 estimators, a maximum depth of 20 and a minimum of 2 samples per leaf. To ensure robustness and increase the generalizability of the models, five-fold cross-validation was employed to assess the predictive performance of each model. By performing feature selection techniques on each cross-validation fold, the potential bias in the prediction performance estimates were mitigated, which enhanced the effective use of the available data. The feature selection and model development is shown in Fig. 4.

Figure 4. Feature selection and model development. (A) Distribution and ranking of optimal radiomic features to predict whether the patients had MLM. (B) Coefficient profiles of the radiomic features in lasso model. Each colored line represents the coefficient path of a radiomic feature. MLM, metachronous liver metastases; lasso, least absolute shrinkage and selection operator; glszm, gray level size zone matrix; LLL, low low low; HHL, high high low.

A total of five models based on machine learning technology were developed in the present study to predict MLM in patients with CRC. Specifically, three radiomics models were constructed using features from HRT2 images of CRC (MRI model), CT images of the liver (CT model) and combined MRI-CT images (merged model). Meanwhile, a clinical model (clinical model) was independently developed using clinical data such as CEA, CA199 and maximum tumor diameter, among others. A merged model combining radiomics information and clinical features (clinical-merged model) was also developed to enhance the predictive performance.

Model evaluation

The predictive performance of the models were evaluated using the area under the receiver operating characteristic (ROC) curve (AUC) over five-fold cross-validation. McNemar's test and Delong test were conducted to compare the classification models. The RF model generated individualized prediction probabilities, which were subsequently subjected to decision curve analysis (DCA) to measure the clinical utility of the Clinical-Merged model for the prediction of MLM. Finally, a visual nomogram was developed to estimate the probability of MLM for each patient based on the clinical-merged model.

Statistical analysis

Statistical analysis of the clinical data was performed using SPSS (version 26.0; IBM Corp.). Differences in categorical characteristics between patients with CRC with and without MLM were compared using Pearson's χ2 test and Fisher's exact test. Continuous variables are presented as the mean ± standard deviation. For normally distributed continuous variables, group comparisons were made using independent samples t-test (for two groups) or one-way ANOVA (for multiple groups). Differences in continuous characteristics between the MLM and non-MLM groups of patients were compared using the Mann-Whitney U test. McNemar's test and Delong test were conducted to compare the classification models. For all statistical analyses, P<0.05 (two-tailed test) was considered to indicate a statistically significant difference.

Results

Clinical characteristics of patients

A total of 157 patients were included in the present study, comprising 104 men with a mean age of 63.99±9.74 years and 53 women with a mean age of 60.53±12.51 years. The patients were divided into MLM (n=67) and non-MLM (n=90) subgroups according to imaging or pathological findings. These data indicate that 67 patients met the diagnostic criteria for MLM through either imaging or pathological confirmation during the 2-year study period, whereas 90 patients showed no clinical evidence of MLM. The baseline characteristics of the patients are summarized in Table I.

Table I. Baseline demographics and clinical characteristics of the patient cohort.

Table I.

Baseline demographics and clinical characteristics of the patient cohort.

Characteristics	MLM (n=67)	Non-MLM (n=90)	P-value
Age, years	62.28±10.91	63.22±10.83	0.59
Sex
Men	48	56	0.22
Women	19	34
CEA, ng/ml	16.21±35.62	11.40±30.44	0.36
CA199, IU/ml	25.34±37.34	28.41±46.70	0.66
Dis, mm	66.61±33.55	69.02±66.61	0.65
Dia, mm	48.71±21.14	51.39±24.76	0.48
T Stage at diagnosis			0.37
T1	0	0
T2	6	15
T3	52	63
T4	9	12
N stage			<0.01a
N0	9	31
N1	21	34
N2	35	1
CRM			0.07
Presence	34	34
Absence	33	56
EMVI			<0.01a
Presence	42	37
Absence	25	53

a P<0.05. MLM, metachronous liver metastases; CEA, carcinoembryonic antigen; CA199, carbohydrate antigen199; Dis, distance from the anus; Dia, maximum tumor diameter; CRM, circumferential resection margin; EMVI, extramural vascular invasion.

Radiomics feature parameters

A total of 1,082 radiomics features were extracted from CT images, which consisted of 14 shape features, 18 first-order intensity features and 1,050 texture features. From the MRI images, 992 radiomics features were extracted, comprising 14 shape features, 18 first-order intensity features and 960 texture features. The texture features included gray level co-occurrence matrix, gray level size zone matrix (GLSZM) and neighboring gray tone difference matrix. Finally, 11 features were selected by the RF algorithm to construct the final radiomics models, with four features from CT and seven features from MRI. The radiomics feature selection results are listed in Table II.

Table II. Radiomics feature selection results.

Table II.

Radiomics feature selection results.

A, CT sequences
Feature name	Classification, feature	Coefficient
Wavelet-LHL_GLSZM_zoneentropy	Texture	−0.0144245
Original_firstorder_entropy	First-order intensity	−0.0143472
Wavelet-HHL_GLSZM_smallarealowgraylevelemphasis	Texture	0.00612166
Wavelet-LLL_GLSZM_graylevelnonuniformitynormalized	Texture	0.03840038
B, MRI sequences
Feature name	Classification, feature	Coefficient
Wavelet-HHH_GLSZM_largeareahighgraylevelemphasis	Texture	−0.0419515
Wavelet-LHL_GLCM_clustershade	Texture	−0.0414815
Wavelet-LHH_NGTDM_coarseness	Texture	−0.0244287
Wavelet-HLH_GLSZM_largeareahighgraylevelemphasis	Texture	−0.0226032
Wavelet-HHH_GLSZM_Largeareaemphasis	Texture	−0.0194299
Log-sigma-0-5-mm-3D_firstorder_skewness	First-order intensity	−0.0015288
Wavelet-HHH_GLSZM_zoneentropy	Texture	0.01368654

[i] GLSZM, gray level size zone matrix; GLCM, gray level co-occurrence matrix; NGTDM, neighboring gray tone difference matrix.

Performance and evaluation of different models

The merged model demonstrated the highest predictive performance of the models analyzed, with sensitivity, specificity and average AUC values of 0.71, 0.72 and 0.82, respectively. The AUC values of the MRI, CT, clinical and clinical-merged models were 0.68, 0.80, 0.62 and 0.75, respectively. The predictive performance of all models were summarized in Table III and Fig. 5. McNemar's test indicated that the radiomics models were significantly different from the clinical model (P=0.01), except for the MRI model (P=0.80). Additionally, according to the Delong test results (z-score=0.52; P=0.60), no significant differences were found between the merged model and the CT model. In terms of clinical utility, the DCA demonstrated that the clinical-merged model had a higher net benefit compared with the other four models. The calibrated probabilities served as inputs to estimate net benefits across varying threshold probabilities. When the threshold probability is <0.38 and >0.93, the use of a radiomics nomogram may offer a higher net benefit compared with the other four models (Fig. 6). To generate quantitative, individualized predictions incorporating key patient characteristics (e.g., EVMI and MR_N status) predictive nomograms were developed (Fig. 7).

Figure 5. ROC-AUC of different models for the prediction of MLM. The models are represented as follows: Clinical, green line; merged, red line; CT, blue line; MRI, orange line; and clinical-merged, purple line. MLM, metachronous liver metastases; ROC, receiver operating characteristic; AUC, area under the curve; merged, the fusion radiomics signature based on features from CT and MRI; clinical-merged, the combined model incorporating clinical and radiological variables together.

Figure 6. Decision curve analysis for the nomogram for the prediction of MLM in all patients. The solid black line indicated the net benefit of assuming that all patients have MLM; the dotted black line indicates the net benefit of the assumption that no patients have MLM and the red line indicates expected net benefit of each patient. At a threshold probability of 0.38, the net benefit of the individualized prediction (red line) equals that of the ‘Treat all’ strategy (black solid line), while at 0.93, it matches the net benefit of the ‘Treat none’ strategy (black dotted line). MLM, metachronous liver metastases.

Figure 7. Nomogram for the prediction of MLM probabilities. A radiomics nomogram, which incorporated imaging histological features and clinical factors. Clinical factors included N stage and EMVI. MLM, metachronous liver metastases; MR_N, N stage of magnetic resonance; EMVI, extramural vascular invasion; glszm, gray level size zone matrix.

Table III. AUCs for all models.

Table III.

AUCs for all models.

Model	AUC	Specificity, %	Sensitivity, %
CT	0.80	75	71
MRI	0.68	57	64
Clinical	0.62	46	68
Merged	0.82	71	72
Clinical-merged	0.75	62	78

[i] AUC, area under the curve; merged, the fusion radiomics signature based on features from CT and MRI; clinical-merged, the combined model incorporating clinical and radiological variables together.

Discussion

Metastases serve a notable role in the mortality of patients with cancer (20). A substantial proportion of patients with CRC present with metastases that are solely or predominantly localized to the liver at the time of diagnosis. Liver metastases that are undetected during initial diagnosis but emerge subsequently in the course of the disease are referred to as MLM (21). In the present study, 67 patients developed liver metastases within 2 years of treatment, with an incidence rate of ~42%, which was higher compared with previous findings reported in the literature (9,22,23). CT is a non-invasive and reliable method of liver assessment and is considered one of the standard imaging modalities for the preoperative detection of liver metastases and postoperative monitoring of patients with CRC (22,23). However, by the time metastases are detected through imaging, it may be too late to effectively intervene. A previous study reported that prophylactic local adjuvant treatment of the liver parenchyma in patients at risk of MLM can effectively reduce the incidence of liver metastases after treatment (24). Therefore, it is important to predict MLM before it manifests. The present study hypothesized that a change in the liver microenvironment may occur before the appearance of imaging manifestations. These changes cannot be detected through conventional examinations; however, the present study aimed to visualize and analyze the changes in the liver microenvironment through radiomics. In patients diagnosed with CRC, physicians assess systemic metastases, including the liver, to determine the appropriate diagnosis and treatment (25). Therefore, the present study aimed to develop a multimodal model using both CT and MRI to capture a comprehensive view of CRC.

In the present study, a total of 11 features were selected to build the radiomics model, which included nine texture-related features (for example, wavelet) and two first-order intensity features. Texture-related features reflect intratumoral heterogeneity, while first-order intensity features extracted through pixel-level analysis enable the quantification of intrinsic image characteristics (8). Wavelet transform extracts multi-frequency and multi-scale image information, which enhances subtle contrast differences between lesions and normal tissues. Wavelet transform is particularly suitable for capturing complex clinical features in tumor images that are difficult to describe with simple visual characteristics, which indicates that images from patients with CRC contain information that is difficult to discern with the naked eye (26). These features provide quantifiable information regarding texture patterns and tissue distribution within the tumor, which may not be easily visually discernible. Among these features, gray level non-uniformity normalized was the best predictor, which measured the similarity of gray-level intensity values in the image. Higher gray level non-uniformity values have been associated with lower similarity, which represent tumor heterogeneity (27). Apart from gray level non-uniformity, small area low gray level emphasis and GLSZM_zoneentropy were also illustrated in the present nomogram. Small area low gray level emphasis has been potentially associated with pathological characteristics such as tumor necrosis and fibrosis (28). Small area low gray level emphasis quantifies the prevalence of small, low-intensity regions within the lesion, which may correspond to necrotic areas or fibrotic tissue deposition. GLSZM_zoneentropy reflects tumor heterogeneity, which serves as a marker of intratumoral complexity. Higher values of GLSZM_zoneentropy indicate greater variability in the size and distribution of homogeneous gray-level zones, which suggest diverse tissue subtype (29). The present study combined traditional observation, radiomics and machine learning to extract multidimensional imaging features of lesions for predictive modeling. These features, whether easily interpretable or not, capture the diversity within tumor regions and reflect the biological variability of CRC. This provides a crucial foundation for building a predictive model of MLM in patients with CRC (30).

By adopting the aforementioned radiological features combined with clinical information, five models were developed to predict MLM in patients with CRC and the performance between these five models were compared. Notably, the merged model exhibited high performance with an AUC value of 0.82, compared with that of the other models. The clinical model, which incorporated parameters such as CEA, CA199, maximum tumor diameter, distance from the anus, T-stage, N-stage, CRM and EMVI, yielded an AUC value of 0.62, whereas the clinical-merged model achieved an AUC value of 0.75. Among these models, the merged model demonstrated optimal predictive performance and McNemar's test results provided statistical validation of its significant diagnostic efficacy (P<0.05). However, the Delong test demonstrated no significant difference between the merged model and the CT model (P=0.60). These findings indicated that the merged model and the CT model may exhibit similar performance, although the merged model exhibited a larger AUC value compared with that of the CT model. Notably, the addition of clinical information failed to improve the performance of the clinical-merged model in terms of AUC. This finding may be attributed to the fact that the clinical data utilized in the present study were obtained through subjective image assessments and the small sample size could have exacerbated these biases. This limitation could be mitigated in future studies involving larger and more diverse cohorts. The present study concluded that radiomics models may outperform clinical models in terms of modelling and indicated that the construction of multimodal models is feasible. However, the clinical significance of this observed difference appears to be more limited than we hypothesized.

Previous radiomics studies have established the prognostic value of CT and MRI features in CRC, which demonstrate potential as predictive biomarkers for MLM (3,31,32). In a multicenter study involving 91 patients with rectal cancer, the radiomics, clinical and merged models achieved AUC values of 0.86, 0.71 and 0.86, respectively (10). However, the incorporation of clinical features failed to significantly enhance the performance of the clinical-merged model, a finding consistent with previous studies and the present study findings (3). This phenomenon may be attributed to the enhanced informational value of radiomics features regarding liver parenchymal characteristics, which exert a notable influence on model performance compared with conventional clinical predictors such as CEA and CA199. Furthermore, a study conducted by Creasy et al (23) emphasized the importance of machine learning analysis of hepatic parenchyma in the venous stage for the identification of patients at a high risk of liver metastases. A number of studies have similarly demonstrated the effectiveness of radiomics for the prediction of MLM (22,33,34). The present study not only expanded the sample size but also extracted comprehensive three-dimensional information from the liver parenchyma. In the present study, the merged model had the best performance according to the results of AUC values, as well as the results of the Delong test and McNemar's test. Therefore, the radiomics model could potentially be used to predict MLM in patients with CRC.

Liang et al (35) extracted features from T2-weighted and venous phase sequence images of 108 patients with rectal cancer, and used support vector machine and logistic regression analysis to develop prediction models. The findings indicated that MRI-based radiomics models derived from baseline rectal images may hold potential for the prediction of MLM (35). Similarly, the present study also tried to construct prediction models using RF algorithms and logistic regression analysis, and a significant performance was achieved using the RF algorithm. Li et al (36) utilized preoperative MRI sequences diffusion-weighted imaging (DWI) and high-definition T2 for predicting MLM in CRC, demonstrating marked efficacy, with the fusion model achieving an impressive AUC value of 0.90. Compared with the aforementioned study, the MRI model generated in the present study performed worse with an AUC value of 0.68. This discrepancy in performance could be attributed to the fact that the present study exclusively selected HRT2 images without contrast, which may have provided less clarity regarding the extent of the lesion compared with enhanced images. Furthermore, the utilization of a single sequence may have limited the amount of information available for analysis, which contributed to the results from the present study.

There were some limitations in the present study. Firstly, the sample size was relatively small, which precludes the application of certain artificial intelligence algorithms, such as deep learning and neural networks. A larger sample size and simultaneous improvement of the algorithms employed will serve a key role in future studies. Secondly, as a single-center retrospective study, external validation with independent datasets is essential to confirm the findings and ensure reproducibility; however, identifying a hospital with consistent examination parameters and scan sequences among the multiple units has proven to be challenging. Standardization of modalities will be a key consideration for future multi-center studies. Thirdly, only the HRT2 sequence of the tumor was selected for analysis, while other sequences such as DWI were not included in the present study. Practical experience demonstrated certain limitations in the precise delineation of DWI sequences and notable interindividual variability was observed. Functional sequences may contain additional information that could be potentially valuable in future studies.

In conclusion, five classification models were developed to predict MLM by analyzing histological imaging and clinical information in patients with CRC. The radiomics models outperformed clinical models alone, which suggested that radiomics may improve the ability to predict MLM in clinical practice. To the best of our knowledge, the present study was the first attempt to integrate two different modalities, CT and MRI, to predict MLM in patients with CRC. Although the diagnostic performance did not exhibit a significant improvement, the present study lays the foundation for future research. Expanding the sample size and incorporating multimodal testing will serve a key role in enhancing the prediction accuracy of MLM in subsequent studies.

Acknowledgements

Not applicable.

Funding

The present study received funding from the Key Laboratory of Functional Molecular Imaging of Tumor and Interventional Diagnosis and Treatment of Shaoxing City, Zhejiang Province Public Welfare Technology Application Social Development Field Project (grant no. LGF20H180008), the Major Program Co-sponsored by Province and Ministry (grant no. WKJ-ZJ-2210) and the General Plan for Medical and Health Research in Zhejiang (grant no. 2020KY323, 2021KY371 and 2023KY1268).

Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

Authors' contributions

JPW, ZNZ, DBS, YNH, WT, HBZ, ZHZ and JHS conceived and designed the present study. JPW, ZNZ, DBS, YNH, WT, HBZ, ZHZ and JHS acquired, analyzed and interpreted the data. JPW drafted the manuscript. JPW, ZNZ, DBS, YNH, WT, HBZ, ZHZ and JHS revised the manuscript. JPW and ZNZ performed the statistical analyses. ZHZ, JHS, YNH and WT obtained funding for the present study. JHS provided the materials for the present study. PW and ZNZ confirm the authenticity of all the raw data. All authors read and approved the final manuscript.

Ethics approval and consent to participate

The present study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The present study was approved by the Ethics Committee of Sir Run Shaw Hospital, Zhenjiang University School of Medicine (approval no. 0465-2022; Zhejiang, China). The need for informed consent was waived by the ethics committee of The Sir Run Shaw Hospital, Zhenjiang University School of Medicine, because of the retrospective nature of the present study.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

References