This retrospective study was approved by the institutional review boards of The First Affiliated Hospital of Soochow University (Approval No. 2024-431) and Nantong Third People’s Hospital (EK 2023025), and the requirement for informed consent was waived. A total of 331 HCC patients were included. The training cohort comprised 228 patients from center I (January 2020 to February 2024: 93 DPHCC and 135 non-DPHCC patients). The external validation cohort consisted of 103 patients from center II (January 2019 to July 2023: 47 DPHCC and 56 non-DPHCC patients).

Inclusion criteria were: (1) Gd-EOB-DTPA-enhanced MRI examination within two weeks before surgery, and (2) pathologically and immunohistochemically confirmed DPHCC or non-DPHCC [Figures 1 and 2]. Exclusion criteria were: (1) any history of anti-tumor therapy prior to surgery, (2) suboptimal image quality precluding reliable interpretation, or (3) incomplete clinical or pathological data. For patients with multiple lesions, only the largest one was analyzed.

All baseline clinical characteristics were collected from medical records, including age, gender, cirrhosis, tumor diameter, tumor number, hepatitis B virus (HBV) status, alpha-fetoprotein (AFP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), and gamma-glutamyl transferase (GGT).

Two experienced pathologists independently reviewed and evaluated all available histological samples. DPHCC pathological diagnostic criteria were as follows: (1) immunohistochemically, at least one hepatocyte marker (such as HepPar-1) showed strong positive expression in more than 15% of the tumor cells, mainly in a diffuse distribution; (2) at least one cholangiocyte marker (such as CK19) and at least one hepatocyte marker were coexpressed in more than 15% of the tumor cells; (3) patients were excluded from the DPHCC group if the tumor contained any independent HCC or ICCA component (regardless of the presence of a transition zone) or if tumor cells did not simultaneously express both HCC and ICCA markers. The 15% cutoff for cholangiocyte marker positivity was adopted from the original definition of Lu et al. [5], this threshold was considered to balance sensitivity and specificity for identifying tumors with dual-phenotype features.

After curative resection, all patients underwent regular surveillance for HCC recurrence. This included serum AFP measurement, liver function tests, and abdominal imaging (ultrasound, contrast-enhanced CT or MRI) within the first postoperative month and subsequently at 3 to 6 month intervals. Recurrence-free survival (RFS) was defined as the interval from the date of surgery to the date of tumor recurrence, death from any cause, or last follow-up. The censored follow-up date was February 2024. For patients lost to follow-up, survival time was calculated from the date of surgery to the date of last available follow-up.

Abdominal MRI was performed using a Siemens Magnetom Skyra 3.0-T MRI scanner (Siemens Healthcare, Munich, Germany) and a Philips Achieva 3.0-T MRI scanner (Philips Healthcare, Best, The Netherlands) with a 16-channel phased-array coil. Patients were placed supine, and the scan covered the region from the diaphragmatic dome to the inferior hepatic edge. Detailed scanning sequences and parameters are provided in the Supplementary Materials. For dynamic imaging, Gd-EOB-DTPA (Bayer, Berlin, Germany) was administered via an antecubital vein (0.1 mL/kg body weight), followed by 20 mL of 0.9% saline at 1 mL/s. The images in arterial phase (AP), portal venous phase (PP) and hepatobiliary phase (HBP) were acquired by performing three-dimensional volume interpolated breath-hold examination (3D VIBE) fat suppression T1WI at 20-30 s, 60-70 s, and 20 min after Gd-EOB-DTPA administration, respectively.

Tumor regions were manually segmented on AP, PP, and HBP images using the ITK-SNAP platform (www.itksnap.org). All segmentations were performed by a radiologist (3 years of experience) and verified by a senior radiologist (13 years of experience). Interobserver reproducibility was assessed in a randomly selected subset of 30 patients. Two radiologists independently delineated volumes of interest (VOIs) covering the entire tumor manually. Each lesion was carefully examined across all phases to accurately define its boundaries and avoid adjacent non-tumor tissue. The VOIs delineated by the senior radiologist were used for subsequent analysis.

All MRI images were interpreted independently by two radiologists with 10 and 13 years of abdominal MRI experience, respectively, who were blinded to histopathology results and other clinical information. Disagreements were resolved by joint review to reach a consensus.

The following radiological features were evaluated according to the definitions in LI-RADS v2018^[18]: (a) rim arterial phase hyperenhancement (rim APHE); (b) non-rim arterial phase hyperenhancement (non-rim APHE); (c) peripheral washout; (d) non-peripheral washout; (e) enhancing capsule; (f) tumor margins (0, smooth margins; 1, non-smooth margins); (g) intratumoral hemorrhage; (h) intratumoral fat; (i) necrosis; (j) peritumoral enhancement in the AP; (k) intratumoral vascularity; (l) delayed enhancement; (m) hepatobiliary phase hypointensity; (n) peritumoral hypointensity in HBP; (o) T2-weighted imaging (T2WI) mild-moderate hyperintensity; (p) targetoid diffusion restriction.

Univariate logistic regression analysis was conducted within the training cohort to screen for statistically significant variables (P < 0.05). Significant variables were entered into a multivariate logistic regression analysis using the forward likelihood ratio method to identify independent predictors, which were then used to construct the radiological prediction model for DPHCC.

Radiomics features (first-order statistics, shape, and texture) were extracted from each VOI using the Pyradiomics package (https://pyradiomics.readthedocs.io/en/latest/). Radiomics analysis was performed with the open-source FeAture Explorer (FAE) software (https://github.com/salan668/FAE), which supported the complete workflow, including feature extraction, feature matrix preprocessing, model development, and visualization.

For DL feature extraction, 3D tumor-region images were cropped based on the VOIs. A 3D Swin Transformer (swin3d_s, PyTorch 2.5.1) was initialized with Kinetics-400 pre-trained weights and fine-tuned on our 3D MRI data using a partial strategy: only the final stage (Stage 4), the PatchMerging layer, the LayerNorm layer, and the classification head were updated. These deeper layers were responsible for capturing task-specific high-level semantic information and were therefore the most beneficial for adaptation to the target domain. For all subsequent deep feature extraction, the resulting fine-tuned weights were used after removing the classification head, with the fine-tuned backbone serving as the feature extractor. To ensure compatibility with the model input requirements, all images were resized to a uniform dimension. Detailed training parameters and the complete fine-tuning code are provided in the Supplementary Materials and on GitHub (https://github.com/doctoryuxing/DPHCC-Swin3d-classification.git).

Feature selection and model building were performed in FAE. To reduce overfitting and improve model robustness, we applied the following feature selection procedure. Initially, Z-score standardization was performed on the feature matrix by subtracting the respective means and scaling with the corresponding standard deviations. Subsequently, we implemented Pearson correlation analysis for feature selection. Pairwise coefficients between features were computed, with a threshold larger than 0.80 triggering random elimination of one feature from correlated pairs. Lastly, RFE (recursive feature elimination) was used to rank the remaining features according to their predictive ability. The feature number range was set from 1 to 30.

Based on the selected features, three machine learning classifiers implemented in FAE software, namely logistic regression (LR), logistic regression-least absolute shrinkage and selection operator (LR-Lasso), and support vector machine (SVM), were used for model construction. Radiomics and DL models were built using radiomics and DL features, respectively. The DLR models incorporated both the radiomics and DL features. In total, 36 models were built based on AP, PP, HBP, and combined-phase (CP) images. Model performance was evaluated using 10-fold cross-validation in the training cohort. The study workflow is shown in Figure 3.

Data processing was performed in R (v4.3.2; https://www.r-project.org/) and SPSS 27.0. Continuous variables were reported as the mean ± standard deviation or median with IQR (interquartile range), while categorical variables were expressed as counts and percentages. Continuous variables were analyzed using either the independent t-test or Mann-Whitney U test based on distribution normality, categorical characteristics were compared with Pearson’s chi-square test. Normality was assessed with the Shapiro-Wilk test. Interobserver agreement for feature extraction was evaluated using the intraclass correlation coefficient (ICC) and categorized as poor (< 0.5), moderate (0.5-0.79), or excellent (≥ 0.8).

Feature correlations were visualized with coefficient matrix heatmaps generated by “ggcorrplot”. Model discrimination was assessed by receiver operating characteristic (ROC) analysis, with AUC values compared by the DeLong test. To assess feature selection stability and overfitting risk, we performed bootstrap resampling (500 iterations) on the training cohort, repeating the entire feature selection pipeline and calculating selection frequency for each feature. Survival probabilities were estimated by the Kaplan-Meier method and compared with the log-rank test. A two-sided P value < 0.05 was considered statistically significant.

The clinical characteristics of patients are presented in Table 1. In the training and external validation cohorts, 40.79% (93/228) and 45.63% (47/103) of patients were confirmed to have DPHCC, respectively [Figure 4]. No statistically significant differences in baseline characteristics were found between the two cohorts except for tumor diameter (P < 0.001). Given that no significant differences were observed in age, sex, cirrhosis, HBV status, ALT, AST, GGT, AFP, tumor diameter, or number between the DPHCC and non-DPHCC subgroups within the training cohort (all P > 0.05), these variables were excluded from the diagnostic model construction.

Radiological characteristics are shown in Table 2. The training and external validation cohorts showed good agreement for all features, with no statistically significant differences (all P > 0.05). In the training cohort, univariate analysis showed that the DPHCC subgroup had higher frequencies of rim APHE (38.7% vs. 5.2%, P < 0.001), peripheral washout (14.0% vs. 3.0%, P = 0.002), and non-smooth tumor margins (50.5% vs. 24.4%, P < 0.001) than the non-DPHCC subgroup. Conversely, non-rim APHE was more common in the non-DPHCC subgroup (88.1% vs. 54.8%, P < 0.001).

All variables significant in univariate analysis were entered into multivariable logistic regression. After adjustment, rim APHE [odds ratio (OR): 4.72, 95% confidence interval (CI): 1.15-19.31] and non-smooth tumor margins (OR: 1.95, 95%CI: 1.01-3.73) remained predictors for DPHCC [Table 3]. In the training cohort, the AUC, sensitivity, and specificity of the radiological model were 0.718 (95%CI, 0.653-0.782), 67.7%, and 67.4%, respectively; in the external validation cohort, the AUC, sensitivity, and specificity of the model were 0.607 (95%CI, 0.511-0.704), 25.5%, and 92.9%, respectively. The ROC curves are shown in Figure 5A and B.

A total of 1,316 radiomics features and 768 DL features were extracted from each VOI per phase. The interobserver ICC was ≥ 0.8, 0.5-0.79, and < 0.5 for 88%, 8%, and 4% of the extracted features, respectively. The mean ICC for VOI delineation was 0.86, indicating satisfactory repeatability. The selected features were suitable for building DPHCC prediction models [Figure 5C-F]. To improve interpretability, we applied class activation mapping (CAM) to visualize the model’s decision patterns and highlight the regions it prioritized. As shown in Figure 6, the concentrated red zones overlapping with blue regions indicated heightened activation, reflecting focused computational attention. Low correlation coefficients in the heatmaps [Figure 7A-D] further supported feature feasibility. The numbers of features selected from each phase and the combined phase are listed in Supplementary Table 1.

LR, LR-Lasso, and SVM classifiers were used to build radiomics, DL, and DLR models. As summarized in Table 4, LR-Lasso performed slightly better overall and was therefore selected for subsequent analyses. The models based on AP and HBP showed unsatisfactory performance in the external validation cohort (AUCs < 0.7), indicating poor diagnostic capability and reproducibility. To reduce redundancy and simplify the models, we therefore focused on the PP and CP models for detailed evaluation. The selected input features are presented in Supplementary Table 2.

Bootstrap stability analysis confirmed that most of the 11 selected features were consistently identified; three features showed high selection frequencies > 70% and six exceeded 25% [Supplementary Table 3], indicating that the core predictive feature set was robust to sampling variability. Although the stability of the low-frequency features was suboptimal, the model as a whole was reliable for risk stratification.

The PP-DLR model achieved the highest predictive performance in both cohorts. In the training cohort, the AUC was 0.864 (95%CI, 0.816-0.911), sensitivity 88.17%, and specificity 70.37%. In the external validation cohort, the AUC was 0.805 (95%CI, 0.722-0.887), sensitivity 76.60%, and specificity 75.00%. The PP-DL model ranked second, with an AUC of 0.801 (95%CI, 0.743-0.859) in the training cohort and an AUC of 0.790 (95%CI, 0.700-0.880) in the external validation cohort. In the external validation cohort, the PP-DLR model showed a significantly higher AUC value than the PP-Radiomics model (P = 0.026), no significant difference was observed between the PP-DL and PP-DLR models (P = 0.784).

Among the CP models, the CP-Radiomics and CP-DLR models had lower AUCs than the CP-DL model, but the differences were not significant (P = 0.281 and 0.219, respectively). The AUC of the CP-DL model was 0.750 (95%CI, 0.684-0.816) in the training cohort and 0.718 (95%CI, 0.620-0.817) in the external validation cohort.

According to the DeLong test, the AUC of the PP-DLR model was significantly higher than those of the CP-Radiomics and CP-DLR models (both P < 0.05). The PP-DLR model had a higher AUC value than the CP-DL model, although the difference was not statistically significant (P = 0.149). The radiological model had significantly lower AUC than those of the PP-DLR and PP-DL models (all P < 0.05), but was comparable to the remaining models (all P > 0.05) [Table 5].

Given the significant difference in tumor diameter between the training and external validation cohorts, a subgroup analysis was performed accordingly. The results showed that tumor diameter had a limited impact on the performance of the PP-DLR model in the external validation cohort (≤ 3 cm: AUC, 0.800 [95%CI: 0.671-0.929]; > 3 cm: AUC, 0.821 [95%CI: 0.709-0.933]; P = 0.805).

As of 28 February 2024, 215 of 228 patients (94.3%) in the training cohort and all 103 patients (100.0%) in the external validation cohort had completed RFS follow-up; 13 patients in the training cohort were lost to follow-up. In both cohorts, DPHCC revealed poorer PFS and earlier recurrence compared with non-DPHCC (both P < 0.05).

In the training cohort, the median RFS was 16.0 (95%CI, 13.6-not evaluable) months for histologically confirmed DPHCC patients and 34.6 (95%CI, 23.8-not evaluable) months for histologically confirmed non-DPHCC patients (P = 0.036) [Figure 8A]. Similarly, the median RFS was 13.9 (95%CI, 11.2-not evaluable) months for patients classified as DPHCC by the PP-DLR model and 35.0 (95%CI, 24.0-not evaluable) months for those classified as non-DPHCC in the training cohort (P < 0.001) [Figure 8B].

In the external validation cohort, the median RFS was 18.0 (95%CI, 12.0-not evaluable) months for histologically confirmed DPHCC and not evaluable (95%CI, 33.0-not evaluable) for histologically confirmed non-DPHCC (P < 0.001) [Figure 8C]; the median RFS was 10.6 (95%CI, 7.1-18.0) months for patients classified as DPHCC by the PP-DLR model and not evaluable (95%CI, 33.3-not evaluable) for those classified as non-DPHCC (P < 0.001) [Figure 8D].

First, the retrospective design introduced inherent selection bias. Second, tumor sizes differed significantly between the training and external validation cohorts. Although subgroup analysis suggested that tumor size had a limited effect on model robustness, these findings are preliminary and need to be validated. Third, the moderate stability of some selected features and the high DPHCC prevalence in our cohort limit generalizability. External validation in larger, multicenter cohorts is needed before clinical translation.

The PP-DLR model may aid in the preoperative identification of DPHCC. DPHCC is associated with poorer recurrence-free survival, and preoperative identification of this subtype may help inform postoperative surveillance strategies.