Patient selection

This study received approval from our Hospital Ethics Committee (2022-CL027-01), and the requirement for informed consent from the enrolled patients was waived. Between January 2017 and July 2023, 287 consecutive HCC patients who underwent preoperative MRI and subsequently received a diagnosis of HCC following hepatectomy were retrospectively analyzed. The exclusion criteria are as follows: (1) transarterial chemoembolization (TACE) or radiofrequency ablation therapy before MRI scanning (n = 6); (2) poor MR image quality or incomplete MRI examinations (n = 7); and (3) the lack of a tumor differentiation report (n = 9). Ultimately, the study involved 265 patients who were randomly classified into either the training or validation cohort with a cross-validation approach at a ratio of 7:3.

Pathological differentiation analysis

The tumor specimens were subjected to hematoxylin and eosin (HE) staining to determine the degree of HCC differentiation by a proficient pathologist with 12 years of experience who was blinded to the preoperative examinations. Based on the classification criteria [19], the tumors were categorized as well-, moderately, or poorly differentiated HCC (pHCC). When HCC tumors displayed various differentiation results, the predominant differentiation determined the final diagnosis. Notably, we classified both moderately and well-differentiated HCC as non-poorly differentiated HCC (npHCC).

MRI protocol

A 3.0-T MRI scanner (Magnetom Verio, Siemens, Germany) was used to perform the MRI scans. The following sequences were conducted: (1) T1WI; (2) T2WI; (3) echo-planar diffusion-weighted imaging (DWI), b values = 0.800 s/mm2; and (4) contrast-enhanced MRI (CEMRI) with an injection of 0.2 mL/kg of Gd-DTPA (Magnevist, Bayer, Germany) at a rate of 1 mL/s for bolus tracking. Three-dimensional (3D) volumetric interpolated breath-hold examination techniques were employed to capture arterial phase (AP, 25–35 s), portal venous phase (PVP, 60–70 s), and delayed phase (DP, 180 s) images. The entire CEMRI procedure lasted approximately 3–4 min. Additional MRI sequence details are shown in Table S1.

Data collection

The baseline data, encompassing patient age, sex, etiology, alpha-fetoprotein (AFP) level, alanine aminotransferase (ALT) level, aspartate aminotransferase (AST) level, total bilirubin (TB) level, prothrombin time (PT), albumin level, Child–Pugh grade, performance status (PS), Barcelona Clinical Liver Cancer (BCLC) stage, HCC number, and differentiation, were collected from the electronic medical record system.

According to the Liver Imaging Reporting and Data System (LI-RADS) criteria (Version 2018) [20], MRI features were evaluated by two radiologists with 7 (H.F.L., radiologist 1) and 12 years (Q.W., radiologist 2) of experience in liver imaging. Both radiologists were blinded to the pathological results, and any inconsistencies were resolved through discussion under the supervision of senior radiologists (W.X., radiologist 3). The following features were collected: (a) major features: non-rim arterial phase hyperenhancement (NAPHE), non-peripheral washout, and enhancing capsules; (b) particular auxiliary features: nodule-in-nodule, mosaic architecture, blood products or higher fat content in HCC; (c) auxiliary features favoring malignancy but not HCC specifically: mild-to-moderate T2 hyperintensity, DP hypointensity, iron sparing in the tumor, and corona enhancement; and (d) baseline features: HCC size, margin, and shape.

Clinical model development

Univariate and multivariate logistic regression methods were subsequently used to determine risk factors for pHCC in the training cohort. These significant factors were then integrated with a logistic regression algorithm to develop a clinical model for identifying pHCC, which was further assessed in the validation cohort.

Imaging preprocessing, ROI segmentation, and peritumoral region dilation

Prior to region of interest (ROI) delineation, the images were resampled at a voxel spacing of 1 × 1 × 1 mm3 and grayscale normalized to compensate for voxel spatial differences and maintain grayscale consistency, respectively. The 3D-ROIs encompassing the entire tumor were manually delineated along the border on each successive transverse slice of the AP, PVP, and DP images by radiologist 1 using the open-source software ITK-SNAP (version 3.6.0, www.itk-snap.org), and all manual delineations were verified by senior radiologist 2. Furthermore, the MR images of 30 HCCs were randomly chosen for resegmentation by both radiologists 1 week later to redraw the ROIs.

As the voxel spacing of the images was resampled to 1 × 1 × 1 mm3, the expanded voxel size was adjusted to the dilated peritumoral region by convolving the ROIs with a 3D box kernel. The peritumoral regions were then acquired with the SimpleITK package in Python software (version 3.6) by dilating the intratumoral 3D-ROIs by 5 mm, 10 mm, and 20 mm in 3D. Notably, areas beyond the liver parenchyma covered by the dilation were manually excluded. Figure 1 shows a representative example of an intratumoral mask and multiscale dilated peritumoral region.

Fig. 1

Examples of masks with different dilation distances on MR images. The red region represents the intratumoral region that was segmented by radiologists. The green-colored ring-like regions indicate the multiple peritumoral regions obtained with dilation

Radiomics feature extraction and dimension reduction

The PyRadiomics package (Version 3.6) was utilized to extract shape (n = 14), first-order (n = 198), texture analysis features (n = 803). Consequently, 1015 features were extracted from each MRI sequence (AP, PVP, and DP) for each region, resulting in the cumulative acquisition of 3045 features from the multiparametric MR images by combining the radiomic features from different regions, including the intratumoral, 5-mm peritumoral (Peri_5mm), 10-mm peritumoral (Peri_10mm), and 20-mm peritumoral (Peri_20mm) regions.

Feature selection criteria and dimension reduction

A three-step procedure was sequentially used to select optimal features [21]. Initially, Spearman’s rank correlation was implemented to eliminate features with a correlation coefficient greater than 0.9. Next, tenfold cross-validation was applied in conjunction with the least absolute shrinkage and selection operator (LASSO) method to select the optimal features with non-zero coefficients, as determined by the optimal penalty parameter. Subsequently, the maximum relevance-minimum redundancy (mRMR) approach was performed to further reduce data dimensionality.

Individual and fusion radiomics model development

The selected robust features were combined according to their respective coefficients to predict pHCC through the support vector machine (SVM) classifier, contributing to the development of individual radiomic models, including intratumoral, Peri_5mm, Peri_10mm, and Peri_20mm models. The prediction probabilities of each peritumoral model were compared to select the most effective model. Subsequently, the features selected from the optimal peritumoral region were integrated with the intratumoral features to design the IntraPeri fusion model.

Statistical analysis

Statistical analysis was performed with the R software (version 4.0). Categorical data are presented as percentages, while continuous variables are expressed as either the mean ± standard deviation or median (interquartile range) following normality testing using the Shapiro–Wilk method. The intraclass correlation coefficient (ICC) was utilized to analyze the variability of radiomic features between and within readers. The Mann–Whitney U or chi-square (χ2) test was employed to ascertain differences between pHCC and npHCC. The performance in predicting pHCC was assessed using the area under the receiver operating characteristic curve (AUC). Moreover, the DeLong method was employed to compare the differences in AUC values among the various models. A comprehensive flowchart illustrating the process from MRI scanning to model development is depicted in Fig. 2.

Fig. 2

Detailed flowchart including MRI scanning, ROI segmentation and peritumoral region dilation, feature extraction and selection, and radiomic model construction and evaluation