Patients
This retrospective study was approved by the Ethic Committee of the Chinese PLA General Hospital (approval no. 2024KY027-KS001), and was conducted in accordance with the 1964 Helsinki Declaration and its later amendments. Informed consent was waived by our Institutional Review Board because of the retrospective nature of our study. From January 2015 to December 2021, all HCC patients were searched in an institutional radiology database. The inclusion criteria consisted of the following: (1) HCCs were diagnosed on imaging criteria outlined in American Association for the Study of Liver Diseases (AASLD) guidelines or by biopsy. (2) a single HCC with a diameter less than 5 cm or multiple HCCs with a maximum diameter less than 3 cm; (3) complete MRI data and clinical examination data before RFA; (4) follow-up for more than 1 year. The exclusion criteria: (1) patients who were previously treated with TACE or systemic chemotherapy; (2) no routine postoperative follow-up; (3) patients with macrovascular involvement of the portal vein and extrahepatic metastases; (4) MRI images with serious image artifacts affecting measurement. A total of 169 HCC patients meeting the inclusion criteria were enrolled in the study. The flow chart is shown in Fig. 1.
RFA treatment
All patients underwent laparoscopic percutaneous RFA guided by ultrasound, utilizing the Cool-tipTM RF Ablation system (Medtronic, MITG, CO, USA). An experienced hepatobiliary interventional surgeon (HB, with 20 years of expertise in interventional procedures) performed the RFA procedures. The procedure used the impedance control mode, utilizing a 17-G cooled-tip electrode with a 2- , 3- or 4-cm exposed tip, accompanied by an internal cold circulation of water (80-140ml/min), and a total ablation time of 12 min. When there is a rapid increase in impedance, the output automatically stops and resumes ablation after a short interval, with an output below 10W. Subsequently, the puncture path was solidified at temperatures ranging from 70 to 90°C.
Image processing
Baseline MRI was performed in all patients using a 3.0-T whole-body MRI system (Discovery 750, GE Healthcare, USA or Skyra, Siemens Medical Solutions, Germany) with phased-array coil. Examinations included standard MRI protocols: T2-weighted imaging (T2WI), DWI, T1-weighted in-phase and out phase imaging, and DCE MRI (Table 1). Contrast media was administered using gadolinium chelate (Magnevist, Bayer) of 0.1 mmol per kilogram of body weight. The dynamic contrast-enhanced images included arterial-phase (AP, 22–25s), portal venous-phase (PVP, 50–70s) and delayed-phase (DP, 180–240s) after the contrast agent administration.
Image analysis
Two independent abdominal radiologists (JT, and WHL, with 4 and 5 years of experience, respectively, in liver imaging), retrospectively analyzed the MR images of HCC prior to RFA treatment. The evaluations were conducted in random order and the observers were blinded to any other results and to each other. The conventional imaging features of HCC evaluated were as follows: (1) number of lesions (single or multiple); (2) lesion size (≤3 cm or >3 cm); (3) lesion location (special or not, special location refers to the special location of the tumor, including the tumor near the diaphragm, gallbladder, gastrointestinal, bile duct, or in the first portal area of the liver or subcapsular liver); (4) tumor capsule (present or absent); (5) HCC with fat content (present or absent); (6) cirrhosis (present or absent).
Then, observers delineated the boundary of HCCs by manually drawing a region of interest (ROI) layer by layer using ITK-SNAP software (version 3.8.0). This process encompassed the whole HCC lesion, while avoiding peripheral blood vessels and satellite nodules. T2WI, DWI, AP, PVP, and DP were obtained for each patient along the boundary of the lesions (Fig. 2). Review and correction were performed by a senior abdominal radiologist (TRL, with 30 years of experience).
Radiomics feature selection
Database management was performed on the United Imaging medical technology intelligent scientific research platform system V1.0 (Beijing United Imaging Intelligent Medical Technology Research Institute). Patient image data and clinical indictors were input, and then 3D volume (3D-VOI) feature extraction, feature reduction, and machine learning model construction were performed on the system. All 169 patients were randomly divided into training cohorts (n=135) and test cohorts (n=34) in a ratio of 8:2. The training cohorts facilitated feature selection and the construction of machine learning models, while the test cohort assessed the performance of the constructed model.
Z-fraction normalization and a B-spline interpolation resampling algorithm were adopted for image preprocessing for each protocol before image feature extraction. Subsequently, the 3D-VOI of T2WI, DWI, AP, PVP, and DP of each HCC extracted 2264 radiomic features.
The intra-class correlation coefficient (ICC) was employed to evaluate the discrepancies among the features derived by observers 1 and 2. The radiomics features with ICC >0.75 were considered to be relatively stable radiomics features. Then, the variance threshold method, select K-best method, and least absolute shrinkage and selection operator (LASSO) regression algorithm were used successively to screen the feature values; finally, the most valuable radiomics features were selected. Among them, K-optimal method selects features with P value <0.05 and other features were removed [13]. Due to its absolute constraint, the LASSO method is capable of shrinking coefficients and setting some coefficients to zero. Therefore, it proves effective for feature reduction and selection.
AI modeling and validation
Three machine-learning classifiers, Support Vector Machine (SVM), Logistic Regression (LR), and Random Forest (RF), were used to construct the predictive model by Supervised Learning. After model training was completed in the training cohort, the internal test cohort was used for model validation, and a comprehensive performance evaluation was performed. The predictive diagnostic efficacy of the model was tested by the receiver operating characteristic (ROC) curve and the confusion matrix, with the indexes included area under curve (AUC) value, sensitivity, specificity, accuracy, recall rate, and F1-score.
Postoperative follow-up
The routine blood examination and hepatorenal function tests were conducted two days to one-week post-operation. Additionally, blood tests were performed to evaluate hepatorenal function and tumor markers, while chest computed tomography and abdominal contrast-enhanced MRI also was underwent 1month post-operation. When the tumor was well controlled, re-examinations were performed once every 2–3 months. During each follow-up visit, patients were re-evaluated in a multidisciplinary treatment setting. Postoperative recurrence was assessed by the senior radiologist (TRL). The presence of HCC recurrence was confirmed if a newly appearing HCC or if distant metastasis was observed. Early tumor recurrence was defined as postoperative recurrence identified within one-year post-surgery [14].
Statistical analysis
Data analysis was performed using commercial software (SPSS Statistics for Windows, version 25.0; IBM, USA). Patient characteristics were compared using the chi-square test or Kruskal-Wallis H test for count data, and t-test, one-way analysis of variance, or rank sum test for measurement data. Receiver operating characteristic (ROC) analysis was utilized to test the diagnostic accuracy of each classifier.
Univariate and multivariate logistic regression analysis was employed to determine the relationship between the relevant clinical data and routine MRI features and early recurrence after RFA in HCC patients in the training test. Variables with P-values <0.10 in the univariate analysis were included in the multivariate analysis. P-values <0.05 was considered of statistically significant difference.