We retrospectively collected data from hepatocellular carcinoma (HCC) patients at two centers from January 2017 to October 2023. Patients from the Second Affiliated Hospital of Zhejiang Chinese Medical (hospital 1), were assigned as the training cohort. Patients from the Qinghai Provincial People's Hospital (hospital 2) were assigned as the validation cohort. Inclusion criteria comprised patients who underwent abdominal enhanced MRI examination within 2 weeks before surgery, received curative resection for HCC, pathologically confirmed as HCC, and had a follow-up period of more than 2 years after surgery or experienced recurrence within less than 2 years of follow-up. Exclusion criteria included patients who underwent interventions such as ablation, chemotherapy, targeted therapy, immunotherapy, or other treatments before surgery, as well as those who died perioperatively. A total of 624 HCC patients were initially included, but due to loss to follow-up, poor imaging quality etc., only 310 patients were ultimately included. A total of 212 patients in the training cohort—68 patients in the early recurrence group and 144 patients in the non-recurrence group. A total of 98 patients in the validation cohort—36 patients in the early recurrence group and 62 patients in the non-recurrence group (Figure 1). The study protocol was approved by the Ethics Committee of the Second Affiliated Hospital of Zhejiang Chinese Medical University and the Qinghai Provincial People's Hospital.

All patients were followed up for at least 2 years after curative resection. In the first month postoperatively, tumor recurrence was screened using serum AFP level assessment, enhanced chest and abdominal CT, or abdominal MRI. Subsequently, follow-up visits were scheduled every 3 months in the first year and every 6 months thereafter. The review endpoint date was October 10, 2023.

The study endpoint was early recurrence, defined as the occurrence of one or more of the following events within 2 years after curative resection surgery: (a) presence of new liver lesions with typical HCC radiological features; (b) histologically confirmed HCC on biopsy or postoperative pathology or non-typical imaging findings suggestive of tumor staining after transarterial chemoembolization (TACE); (c) extrahepatic metastasis confirmed by typical radiological features or histological analysis.

Query and record relevant clinical information of patients with liver cancer before surgery from our hospital's electronic medical record system, including patient gender, age, presence of liver cirrhosis, AFP (alpha-fetoprotein) level, ALT (alanine aminotransferase) level, AST (aspartate aminotransferase) level, ALP (alkaline phosphatase) level, TBIL (total bilirubin) level, and ALB (albumin) level.

The MRI scanners used in this study were the Siemens Magnetom Avanto 1.5 T (hospital 1) and the Siemens Skyra 3.0 T (hospital 2). All patients were required to fast for at least 4 h before the examination and undergo breath-holding training prior to scanning. The scanning range extended from the top of the diaphragm to the lower edge of the liver. The scanning parameters (hospital 1) were set as follows: (1) T1WI: TR = 5.6 ms, TE = 2.1 ms, slice thickness = 3.5 mm, interslice gap = 5 mm; (2) T2WI: TR = 3,100 ms, TE = 76 ms, slice thickness = 4.5 mm, interslice gap = 5 mm; (3) DWI: TR = 6,900 ms, TE = 59 ms, slice thickness = 4.5 mm, interslice gap = 5 mm, with b-values of 50 and 1000 s/mm^2; (4) DCE-MRI TR = 3.8 ms, TE = 1.5 ms, slice thickness = 3 mm, interslice gap = 0 mm; And the scanning parameters (hospital 2) were set as follows: (1) T1WI: TR = 3.5 ms, TE = 1.5 ms, slice thickness = 5 mm, interslice gap = 1 mm; (2) T2WI: TR = 2,800 ms, TE = 100 ms, slice thickness = 5 mm, interslice gap = 1 mm; (3) DWI: TR = 8,500 ms, TE = 80 ms, slice thickness = 5 mm, interslice gap = 1 mm, with b-values of 400 and 1000 s/mm^2; (4) DCE-MRI TR = 4.5 ms, TE = 2 ms, slice thickness = 3 mm, interslice gap = 0.5 mm. Liver multiphase dynamic contrast-enhanced imaging: Gd-DTPA (gadolinium-diethylenetriamine pentaacetic acid) was used as the contrast agent in all cases. It was administered intravenously at a dose of 0.1 mmol/kg and an injection rate of 2 ml/s, followed by a 20 ml flush of normal saline. Contrast-enhanced scans were acquired during the arterial phase, portal venous phase, and delayed phase.

Qualitative analysis of MR image features was independently conducted by two abdominal radiologists with 10 and 20 years of diagnostic experience in abdominal imaging, respectively. The radiologists were blinded to radiological and pathological reports. Both physicians observed the features twice, with the second observation occurring 2 weeks after the first. On a per-patient basis, the reviewers evaluated 10 imaging features that have been reported to describe HCC (16–18), including: Tumor number; Tumor size, defined as the maximum diameter on axial delayed-phase images; Arterial phase rapid enhancement (Nonrim-like enhancement of the tumor in the arterial phase unequivocally greater in whole or in part than the liver); portal phase rapid clearance (Nonperipheral visually assessed temporal reduction in the enhancement of the tumor in whole or in part relative to composite liver tissue in the portal venous phase or delayed phase); Peritumoral enhancement (refers to the wedge-shaped or irregular enhancement around the tumor in the late arterial phase or early portal venous phase); Intratumoral hemorrhage (Intralesional or perilesional hemorrhage in the absence of biopsy, trauma, or intervention); Intratumoral necrosis (Presence of nonenhancing area in a solid mass, not attributable to a cystic component, prior treatment, or intralesional hemorrhage); Tumor margin (whether the tumor margin is smooth and uniformly enhanced in the portal phase or delayed phase); Pseudo capsule (A fibrous reactive band surrounding the tumor, visible across multiple imaging phases, that resembles a “capsule” in both morphology and enhancement characteristics, yet is not a true anatomical capsule); macrovascular invasion (unequivocally enhancing soft tissue in vein).

Quantitative data was presented as the mean ± standard deviation (M ± SD) when the data distribution was normal, or median and interquartile range (IQR) suitable for the data of abnormal distributions, and enumeration data was recorded as frequency (percentages), as appropriate. The chi-square or Fisher's exact test was used for categorical variables and the Student t or Mann–Whitney U-test for continuous variables in univariate analysis. A value of P relaxed to less than 0.05 was considered to indicate a significant difference in univariate analysis. Include variables with P < 0.05 from univariate analysis in multivariate analysis. Any variable with a P value ≤ 0.05 was retained in the final model. To derive a simple-to-compute and optimal score, we converted regression coefficients to weighted scores by dividing each regression coefficient by one-half of the smallest coefficient and rounding to the nearest integer or taking the integer part. For each patient, the individual score corresponding to the predictors was summed together to produce an overall score. ROC curve and calibration curve are used to evaluate the performance of the model. P value of < 0.05 was considered statistically significant. All statistical analysis were performed using SPSS software (IBM SPSS Statistics Version 25.0; Origin 2019b).

There was no significant difference in clinical characteristics between the training cohort and the validation cohort (Table 1). In the training cohort, no statistically significant differences were observed in demographic characteristics and laboratory parameters between the early recurrence group and the non-recurrence group after liver cancer resection (P > 0.05, Table 2).

MRI imaging features that showed statistically significant differences in univariate analysis, including Tumor number, Tumor margin, Peritumoral enhancement, and macrovascular invasion (Table 2). Further multivariable logistic regression revealed that Tumor number, Tumor margin, Peritumoral enhancement, and macrovascular invasion are independent predictors for early postoperative recurrence after HCC resection (Table 3).

In the univariate analysis of the training cohort, variables with P < 0.1 were selected for collinearity diagnostics, which demonstrated no significant multicollinearity (VIF < 5). Subsequently, through multivariate logistic regression analysis, tumor number, tumor margin, peritumoral enhancement, and macrovascular invasion showed statistically significant differences (P < 0.05). A traditional predictive model was constructed using these four features for preoperative prediction. The Hosmer–Lemeshow goodness-of-fit test showed good calibration of this predictive model (p = 0.427), and the model exhibited good performance with an AUC of 0.876 (95% CI, 0.817–0.923) (Figure 2).

A weighted score was assigned to each independent predictor associated with early recurrence after HCC resection as follows: The preoperative model: 2 points, tumor number; 3 points, tumor margin; 2 points, peritumoral enhancement; 2 points, macrovascular invasion (Table 3). For each patient, the individual scores that correspond to all the independent predictors were summed together to produce an overall score, which was named as score model. In our study, the score range for the preoperative model is 0–9 points, with a median score of 2 points. The higher the score, the greater the probability of early recurrence after liver cancer resection (Figure 3). The Hosmer-Lemeshow goodness-of-fit test indicated that the scoring model had good calibration (p = 0.790). Additionally, we plotted the calibration curve for the scoring model, which demonstrated good concordance (Figure 4). The AUC of this discriminative scoring system, as measured by the receiver operating characteristic (ROC) curve analysis, was 0.873 (95% CI, 0.803–0.943) (Figure 2). Using 3.5 as the cutoff value, the model showed high predictive efficiency, with a sensitivity of 83.9%, specificity of 77.8%, and accuracy of 80.0% (Table 4). A comparison of the ROC curves showed no significant difference between the main predictive model and the scoring model (p = 0.132), indicating that the scoring model effectively utilized the predictive model's values and could accurately predict early recurrence after liver cancer resection.

To validate the predictive performance of the scoring model, we conducted external validation, and the results of the scoring system were satisfactory. The Hosmer-Lemeshow goodness-of-fit test demonstrated good calibration (p = 0.694). In the validation cohort, the AUC of the scoring system was 0.847 (95% CI, 0.791–0.921). At a cutoff value of 3.5, the model's sensitivity, specificity, and accuracy were 81.3%, 73.8%, and 77.2%, respectively (Table 4), similar to the results in the training cohort.

To apply the score system conveniently in clinical work, we further divided scores into 4 separate groups: preoperative score: ≥0 point and ≤2 points, >2 points and ≤4 points, >4 points and ≤6 points and >6 points and ≤9 points. The predictive probability of early recurrence after HCC resection increases as stage increase of the score (Table 5).