Interpretable prediction of macrotrabecular-massive HCC and recurrence-free survival by integrating MRI LI-RADS features with deep learning habitat radiomics Authors: Mengwei Zhang, Qian Wu, Mo Zhu, Tao Zhang, Chunyan Gu, Wenhao Gu, Mingzhan Du, Hailin Shen, Chunhong Hu, Yan Yang, Ximing Wang, Yixing Yu Published: 2026 ABSTRACT AbstractObjectiveMacrotrabecular-massive (MTM+) hepatocellular carcinoma (HCC) is associated with poor prognosis and early recurrence. We developed and validated a nomogram integrating magnetic resonance imaging LI-RADS features with deep learning (DL) habitat radiomics for preoperative prediction of MTM + HCC and stratifying patients according to recurrence-free survival (RFS).Materials and methodsIn this retrospective multicenter study, 607 patients with early-stage HCC who underwent curative-intent surgical resection (mean age ± standard deviation, 59.6 ± 10.6 years; 474 males, 133 females) were divided into the training (n = 304), internal validation (n = 131), and external (n = 172) test sets. Liver Imaging Reporting and Data System (LI-RADS) features, LI-RADS categorization, and clinical features were analyzed. Finally, a nomogram integrating LI-RADS features, habitat radiomics, and DL features was developed using multivariate logistic regression. Multivariable Cox regression analyses were performed to identify independent prognostic factors.ResultsAt multivariable analysis, habitat radiomics score, DL score, alpha-fetoprotein (AFP), alanine transferase (ALT), and fat mass were independent predictors of MTM + HCC. The DL habitat radiomics nomogram demonstrated powerful performance with areas under the receiver operating characteristic curve (AUCs) of 0.897 (95% confidence interval [CI]: 0.854–0.940), 0.757 (95% CI: 0.674–0.827), and 0.858 (95% CI: 0.797–0.906) in the training, internal validation, and external test sets, respectively. After multivariable analysis, AFP (p < 0.001) and nomogram-predicted MTM state (p = 0.039) resulted as independent prognostic factors.ConclusionThe nomogram integrating LI-RADS features, habitat radiomics, and DL features accurately identified patients with MTM + HCC and stratified patients by RFS.Key Points Question Preoperative integrative methods for accurate prediction of MTM + HCC and postoperative recurrence risk stratification remain insufficient.Findings The nomogram integrating LI-RADS features and DL habitat radiomics effectively diagnosed MTM + HCC and showed superior diagnostic performance compared to the Clinical-radiologic model.Relevance statement The nomogram provides a robust, non-invasive tool for preoperative identification of MTM + HCC and recurrence risk stratification, facilitating surgical planning and postoperative surveillance strategies. Graphical Abstract Keywords: Carcinoma (hepatocellular), Deep learning, Magnetic resonance imaging, Nomograms, Radiomics Background Primary hepatocellular carcinoma (HCC) is one of the most common malignant tumors of the liver, and has a high recurrence rate after surgical resection [1, 2]. Previous studies have demonstrated that angiogenesis and intratumoral heterogeneity are key features of HCC and can be evaluated noninvasively [3, 4]. Macrotrabecular-massive (MTM + ) HCC, a distinct histological subtype of HCC, has garnered considerable attention due to its association with poor prognosis and early recurrence [5–7]. A recent study has identified Endothelial-Specific Molecule 1 as a potential reliable biomarker for MTM + HCC [8]. However, immunohistochemical detection requires pathological specimens, highlighting the important role of preoperative noninvasive diagnosis. Several diagnostic models based on radiomics features have demonstrated good performance in predicting MTM + HCC [9, 10]. Traditional radiomics models primarily focused on features extracted from the entire tumor, often overlooking its spatial heterogeneity. In contrast, habitat radiomics analyzes imaging features derived from specific habitats within the tumor, offering more nuanced insights than conventional whole-tumor analysis [11, 12]. Xie et al [13] established a prognostic model using habitat radiomics features derived from specific subregions in esophageal tumors, which achieved superior predictive performance compared to a whole-tumor-based radiomics model. Previous studies have shown that HCC, particularly MTM + HCC, is marked by extensive heterogeneity in its imaging characteristics [14–16]. Given this heterogeneity, habitat radiomics may improve performance in identifying MTM + HCC. Deep learning (DL) models, which leverage large datasets to select effective imaging features, have been successfully applied for the preoperative prediction of vessels encapsulating tumor clusters and microvascular invasion in HCC [17–19], demonstrating high predictive performance. Li et al [20] developed a DL radiomics model using dual-energy computed tomography (CT), achieving an area under the receiver operating characteristic curve (AUC) of 0.89 (95% confidence interval [CI]: 0.81–0.97) for predicting MTM + HCC in the external test set. To our knowledge, no studies have simultaneously utilized both habitat radiomics and DL features to predict MTM + HCC. Our study aimed to develop and validate a DL habitat radiomics fusion nomogram based on gadolinium-ethoxybenzyl-diethylenetriamine pentaacetic acid (Gd-EOB-DTPA)-enhanced MRI that integrated clinical, radiologic, habitat radiomics, and DL features to predict MTM + HCC and stratify patients by recurrence‑free survival (RFS) risk. Materials and methods Ethics approval This multicenter study was approved by the institutional review boards of all three participating hospitals (Institute 1: the First Affiliated Hospital of Soochow University, No. 2024431; Institute 2: the Nantong Third People’s Hospital, No. EK2023025; Institute 3: the First People’s Hospital of Taicang, No. 2022-ky-203), and informed consent was waived due to its retrospective nature. Patients and clinical data This retrospective study was approved by the institutional review boards of all participating medical centers, with a waiver of the requirement for informed consent. The study included 739 patients with pathologically confirmed HCC between June 2015 and January 2024 from three centers. Patients were included if they (1) underwent Gd-EOB-DTPA enhanced MRI within 14 days before the surgery; (2) underwent partial hepatectomy; and (3) had postoperative histology results indicating whether the tumor was MTM + HCC. Patients were excluded if they (1) underwent local-regional therapy before surgery; (2) had incomplete clinical data or poor image quality. (3) exhibited distant metastasis or other malignant diseases. Finally, a total of 607 patients with early-stage HCC who underwent curative-intent surgical resection were included; 435 patients from center 1 were randomly assigned to a training set (n = 304) and an internal validation set (n = 131) with a ratio of 7:3, and patients from centers 2 and 3 (n = 172) were used as the external test set. The flowchart of patient recruitment is shown in Fig. 1. If multiple lesions were present in the liver, the largest lesion was selected for the study. a Flowchart of patient selection. b Flowchart of patients enrolled for recurrence-free survival analysis, derived from the final cohort in a. Gd-EOB-DTPA, Gadolinium-ethoxybenzyl-diethylenetriamine pentaacetic acid, MTM + HCC, Macrotrabecular-massive hepatocellular carcinoma Clinical and laboratory data for all patients were retrospectively collected from electronic medical records, including sex, age, viral hepatitis status, cirrhosis, tumor number, maximum tumor diameter, and levels of alpha-fetoprotein (AFP), aspartate transaminase (AST), alanine transferase (ALT), and γ-glutamyl transferase (GGT). Histopathology analysis All histologic slides were independently reviewed by two senior pathologists (C.G. and M.D., both with 8 years of experience). According to the 2019 World Health Organization classification [5], MTM structure was defined as the presence of tumor cell layers exceeding 6 layers. When more than 50% of the tumor volume exhibited MTM structure, it was classified as MTM + HCC, and those with less than 50% MTM structure were categorized as MTM- HCC. The interobserver agreement for MTM classification was excellent (κ = 0.92) and was assessed in the entire cohort of 607 patients from all three centers. MRI acquisition MRI was performed using three 3-T systems: MAGNETOM Verio (Siemens Healthcare, Erlangen, Germany), Achieva (Philips Medical Systems, Best, The Netherlands), and Discovery MR750 (GE Healthcare, Chicago, IL, USA), each equipped with 16-channel phased array torso coils. Gd-EOB-DTPA (Primovist; Bayer Schering Pharma, Leverkusen, Germany) was injected using a high-pressure syringe at a flow rate of 1–2 mL/s, with a total dose of 0.025 mmol/kg. Dynamic phases were acquired after contrast injection using voxel-interpolated breath-hold fat-suppressed T1-weighted imaging. Arterial phase (AP), portal venous phase (PP), and hepatobiliary phase (HBP) images were obtained at 30 s, 60 s, and 20 min after contrast injection, respectively. Specific parameters for the different scanners are detailed in Supplementary Table S1. MRI feature analysis All MRI data were retrospectively analyzed by two radiologists (M.Z. and Q.W., with 3 and 5 years of experience, respectively). The following qualitative imaging features were evaluated primarily based on Liver Imaging Reporting and Data System (LI-RADS) and previous studies [21–23]: infiltrative appearance, fat in mass, blood product in mass, enhancement pattern (nonrim arterial phase hyperenhancement [APHE] and rim APHE), washout pattern (nonperipheral washout and peripheral washout), type of capsule (enhancing capsule and non-enhancing capsule), AP peritumoural enhancement, intratumoural artery, substantial necrosis (> 20%, > 50%), satellite nodules, central scar, peritumoral hypointensity in the HBP, gadolinium-enhanced clouding, LI-RADS category. Inter-reader agreement for the above LI-RADS-related imaging features and LI-RADS category, assessed in the entire cohort of 607 patients from all three centers, was excellent (κ range: 0.86–0.91). Details of features are described in Supplementary Material Appendix S1. Automatic segmentation using nnU‑Net Three-dimensional volume-of-interest regions of 353 HCC lesions in three data sets were manually contoured on AP, PP, and HBP images by two junior radiologists (M.Z. and Q.W., with 3 and 5 years of experience) in consensus using 3D-slicer 5.6.2 (https://www.slicer.org/). The nnU-Net framework (https://github.com/MIC-DKFZ/nnUNet) was employed for automatic segmentation. The 353 HCC lesions were randomly split into training (n = 282) and validation sets (n = 71). The model was trained using the default three-dimensional full-resolution configuration of nnU-Net. Detailed procedures and performance metrics are provided in the Supplementary material (Appendix S2 and Fig. S1). The remaining 254 lesions underwent automatic segmentation using the nnU-Net framework. A senior radiologist (Y.Y., with 12 years of experience) reviewed both the manual segmentations from the two junior radiologists and the model-predicted masks, manually correcting any inaccurate masks to establish the ground truth segmentations. Ultimately, 607 ground truth segmentations verified by the senior radiologist were used for subsequent analyses. Habitat radiomics feature extraction K-means clustering (k = 2–8) was applied to radiomic features, with the Calinski–Harabasz index used to evaluate cluster validity. Based on the highest Calinski–Harabasz index, the optimal number of clusters was determined to be 3 (Supplementary Fig. S2). The derived cluster centroids were then applied unchanged to the internal validation and external test sets to generate three habitat subregion maps, serving as imaging biomarkers for quantifying intratumor heterogeneity. Habitat radiomics features were extracted using PyRadiomics software version 3.0.1 (http://www.radiomics.io/pyradiomics.html). Features were derived from the original, Laplacian-of-Gaussian (with sigma values ranging from 2.0 to 5.0 mm), and wavelet-transformed (eight directional decompositions) images. The extracted feature classes included shape, first-order statistics, gray-level co-occurrence matrix–GLCM, gray-level run-length matrix–GLRLM, gray-level size zone matrix–GLSZM, and gray-level dependence matrix–GLDM. Before feature extraction, all images were resampled to an isotropic voxel spacing of 1 × 1 × 1 mm³ to ensure scale invariance. To compute texture features, image intensities were discretized using a fixed bin width of 5. For each cluster, 1,132 habitat radiomics features were extracted, totaling 3,396 features for a single phase per patient. DL feature extraction A pretrained (Kinetics dataset-based) 3D Swin Transformer model from the “torchvision” library of the DL framework PyTorch v2.5.1 (https://github.com/pytorch/pytorch) was used to develop our DL models using transfer learning. The code for the 3D Swin Transformer model is available on GitHub (https://github.com/doctoryuxing/MTM-HCC-Swin3d-classification.git). Before model input, all images underwent intensity normalization to scale pixel values to the [0,1] range to reduce scanner-related variations. Additionally, the Swin Transformer architecture incorporates Layer Normalization, which provides feature-wise normalization during training and enhances robustness to input variations. To prevent overfitting during training, we implemented several proven techniques, including advanced data augmentation and scheduled learning rate decay. The augmentation pipeline included transformations such as Resized, ScaleIntensityRanged, RandFlipd, and RandAffined. We used a 16-batch size and an initial learning rate of 0.0001, with the learning rate decaying by a factor of 0.1 every 7 epochs. To handle class imbalance, we applied focal loss and the Adam optimizer, which was selected for its automatic learning rate adjustments. The training process lasted between 200 and 500 epochs. After completion of training, the final model weights were saved and used for subsequent deep feature extraction. The training and validation learning curves of the AP are provided in Supplementary Fig. S3, demonstrating stable convergence without overfitting. The process is presented in Fig. 2. Workflow of the study, including data collection (a), feature extraction and selection (b), model construction (c), and model assessment (d). AFP, Alpha-fetoprotein; ALT, Alanine aminotransferase; DL, Deep learning; MTM, Macrotrabecular-massive; VOI, Volume-of-interest Construction of models In the training set, univariable analysis and multivariable stepwise logistic regression analysis with Akaike information criterion were used to determine the clinical and radiologic features to include in the clinical-radiologic (CR) model. Habitat radiomics features from three phases (AP, PP, and HBP) were integrated to generate the habitat radiomics combined-phase (CP) features. Similarly, DL features extracted from the same three phases were fused to create the DL CP features. Habitat radiomics and DL features were processed and selected by FeAture Explorer V 0.5.12 (https://github.com/salan668/FAE). The detailed procedure is listed in the Appendix S3. After feature normalization and selection, habitat radiomics models were built with the support-vector-machine (SVM) classifier, whereas DL models were constructed using the autoencoder classifier. The best-performing habitat radiomics and DL models were selected according to their respective internal validation set AUCs. Habitat radiomics score and DL score for MTM + HCC were calculated based on the mean predicted probabilities from their respective best-performing models. To enhance the interpretability of the DL models, we generated class activation maps to visualize the key regions identified by the models. The class activation map utilized color mapping to display the importance of different tumor regions in the classification task. DL habitat radiomics fusion nomogram construction All variables with p < 0.05 (including clinical, radiologic, habitat radiomics score, and DL score) were entered into a multivariable backward logistic regression to identify independent predictors of MTM + HCC. Finally, all selected variables were used to build the DL habitat radiomics fusion model using the Logistic regression method. The DL habitat radiomics fusion nomogram was generated for visualization of the model. Calibration curves and confusion matrices were used to evaluate the performance of the nomogram. Decision curve analysis was applied to assess the clinical benefits of the nomogram. Model interpretability was further analyzed using Shapley additive explanations (SHAP). Cox regression analysis for RFS Patients underwent postoperative clinical and radiologic follow-up at intervals of 3–6 months (Fig. 1). RFS was calculated from the date of surgery to the radiographic detection of tumor recurrence or death, or until the last follow-up if no recurrence was observed. Follow-up continued until December 30, 2024. On the basis of the training set, clinical features and nomogram-predicted MTM state (predicted MTM+ and predicted MTM−) were entered into univariable and multivariable Cox regression analysis. Independent predictors (p < 0.05) were then integrated to construct a prognostic model for RFS. The model performance was assessed by the C-index. Statistical analysis Statistical analyses were conducted using SPSS 26.0, R software 4.4.1, and Python 3.12. Categoric variables are expressed as numbers and percentages, continuous variables are expressed as the mean ± standard deviation or median with interquartile range. The Chi-Square test was used to compare frequencies of categorical variables between MTM+ HCCs and MTM- HCCs. The Kolmogorov–Smirnov test was used to assess the normality of continuous variables. The independent-samples t-test or the Mann–Whitney U-test was performed for continuous variables. Spearman analysis was used to evaluate the correlation between features. The AUCs of different models were compared by the DeLong test. Kaplan–Meier analysis with the log-rank test was used to compare the prognosis rate. For the pairwise comparisons among the habitat radiomics models and DL models, the Bonferroni correction was applied by adjusting the significance threshold to p < 0.0083 to control the family-wise error rate. For all other analyses, a two-tailed p < 0.05 was considered indicative of a statistically significant difference. Receiver operating characteristic curves were drawn with MedCalc 20.022 (https://www.medcalc.org/). R packages including “survival”, “survminer” were used for survival analysis and visualization. The “shap” in Python was used to generate SHAP of the model. Results Clinical and radiologic characteristics of patients The clinical and radiologic characteristics of patients are presented in Tables 1 and 2. A total of 607 patients (474 males and 133 females; aged 59.6 ± 10.6 years [mean ± standard deviation]) were included in this study, with 151 MTM + HCC (122 from center 1, 29 from centers 2 and 3). In the training set, patients diagnosed with MTM + HCC were younger and exhibited larger maximum tumor diameters compared to MTM- HCC patients (p = 0.028 and < 0.001, respectively). Among other variables, ALT (p = 0.008), AST (p < 0.001), GGT (p = 0.013), and AFP (p < 0.001) were found to be significantly higher in the MTM+ group than in the MTM- HCC group. There were no significant differences in the remaining baseline characteristics between the two groups. Table 1Patient characteristics in the training, internal validation, and external test setsCharacteristicsTraining set (n = 304)Internal validation set (n = 131)External test set (n = 172)p-valueap-valuebTotalPositive MTM (n = 88)Negative MTM (n = 216)p-valueSex0.1000.4620.120 Male237 (78.0)74 (84.1)163 (75.5)108 (82.4)129 (75.0) Female67 (22.0)14 (15.9)53 (24.5)23 (17.6)43 (25.0)Age60.05 ± 11.2157.84 ± 10.1260.94 ± 11.530.02860.13 ± 10.2458.49 ± 9.650.1130.301Viral hepatitis0.556< 0.0010.019 No82 (27.0)20 (22.7)62 (28.7)29 (22.1)20 (11.6) HBV212 (69.7)66 (75.0)146 (67.6)98 (74.8)151 (87.8) HCV7 (2.3)1 (1.1)6 (2.8)2 (1.5)1 (0.6) HBV + HCV co-infection3 (1.0)1 (1.1)2 (0.9)2 (1.5)0 (0.0)Cirrhosis0.156< 0.0010.048 Yes143 (47.0)47 (53.4)96 (44.4)69 (52.7)110 (64.0) No161 (53.0)41 (46.6)120 (55.6)62 (47.3)62 (36.0)ALT(%)0.0080.0690.381 > 50 U/L61 (20.1)26 (29.5)35 (16.2)30 (22.9)47 (27.3) ≤ 50 U/L243 (79.9)62 (70.5)181 (83.8)101 (77.1)125 (72.7)AST(%)< 0.0010.0100.100 > 40 U/L90 (29.6)39 (44.3)51 (23.6)42 (32.1)71 (41.3) ≤ 40 U/L214 (70.4)49 (55.7)165 (76.4)89 (67.9)101 (58.7)GGT (%)0.0130.9330.220 > 60 U/L119 (39.1)44 (50.0)75 (34.7)61 (46.6)68 (39.5) ≤ 60 U/L185 (60.9)44 (50.0)141 (65.3)70 (53.4)104 (60.5)AFP(%)< 0.0010.0570.262 > 100 ng/mL124 (40.8)53 (60.2)71 (32.9)50 (38.2)55 (32.0) ≤ 100 ng/mL180 (59.2)35 (39.8)145 (67.1)81 (61.8)117 (68.0)Tumor number0.5900.0130.088 ≥ 290 (29.6)28 (31.8)62 (28.7)36 (27.5)33 (19.2) 1214 (70.4)60 (68.2)154 (71.3)95 (72.5)139 (80.8)Maximum diameter (cm)4.10 (2.50, 7.18)6.30 (3.52, 9.78)3.60 (2.40, 5.70)< 0.0013.90 (2.50,6.30)3.00 (1.90, 4.58)< 0.0010.001Data are numbers of patients, with percentages in parenthesesAFP Alpha-fetoprotein, ALT Alanine aminotransferase, AST Aspartate aminotransferase, GGT Gamma-glutamyl transferase, MTM + HCC Macrotrabecular-massive hepatocellular carcinomap-value < 0.05 indicates a significant difference between MTM + HCC and MTM- HCC in the training seta p-value < 0.05 indicates a significant difference between the training and external test setb p-value < 0.05 indicates a significant difference between the internal validation and external test set Patient characteristics in the training, internal validation, and external test sets Data are numbers of patients, with percentages in parentheses AFP Alpha-fetoprotein, ALT Alanine aminotransferase, AST Aspartate aminotransferase, GGT Gamma-glutamyl transferase, MTM + HCC Macrotrabecular-massive hepatocellular carcinoma p-value < 0.05 indicates a significant difference between MTM + HCC and MTM- HCC in the training set a p-value < 0.05 indicates a significant difference between the training and external test set b p-value < 0.05 indicates a significant difference between the internal validation and external test set Table 2Radiologic characteristics in the training, internal validation, and external test setsCharacteristicsTraining set (n = 304)Internal validation set (n = 131)External test set (n = 172)p-valueap-valuebTotalMTM + HCC (n = 88)MTM- HCC (n = 216)p-valueInfiltrative appearance99 (32.6)41 (46.6)58 (26.9)0.00146 (35.1)60 (34.9)0.6060.967Blood product in mass66 (21.8)31 (35.2)35 (16.3)< 0.00127 (20.6)24 (14.0)0.0360.125Fat in mass71 (23.4)9 (10.2)62 (28.7)0.00121 (16.0)36 (20.9)0.5430.280Enhancement pattern0.0660.5080.330 Nonrim APHE255 (83.9)71 (80.7)184 (85.2)109 (83.2)137 (79.7) Rim APHE31 (10.2)14 (15.9)17 (7.9)10 (7.6)22 (12.8)Washout0.0170.0610.005 Nonperipheral washout208 (68.4)61 (69.3)147 (68.1)83 (63.4)128 (74.4) Peripheral washout18 (5.9)10 (11.4)8 (3.7)6 (4.6)15 (8.7)Capsule0.0480.0090.548 Non-enhancing capsule41 (13.5)15 (17.0)26 (12.0)10 (7.6)8 (4.7) Enhancing capsule137 (45.1)46 (52.3)91 (42.1)63 (48.1)84 (48.8)AP peritumoural enhancement44 (14.5)15 (17.0)29 (13.4)0.41624 (18.3)19 (11.0)0.2890.072Intratumoural artery118 (38.8)50 (56.8)68 (31.5)< 0.00143 (32.8)42 (24.4)0.0010.107Substantial necrosis< 0.0010.0730.902 > 20%91 (29.9)34 (38.6)57 (26.4)31 (23.7)37 (21.5) > 50%29 (9.5)15 (17.0)14 (6.5)10 (7.6)13 (7.6)Satellite nodules40 (13.2)16 (18.2)24 (11.1)0.09822 (16.8)6 (3.5)0.001< 0.001Central scar44 (14.5)19 (21.6)25 (11.6)0.02417 (13.0)31 (18.0)0.3070.233Peritumoural hypointensity in the HBP45 (14.8)24 (27.3)21 (9.7)< 0.00121 (16.0)17 (9.9)0.1260.109Gadolinium-enhanced clouding18 (5.9)4 (4.5)14 (6.5)0.5179 (6.9)19 (11.0)0.0450.214LI-RADS0.0220.1010.408 LR-320 (6.6)2 (2.3)18 (8.3)12 (9.2)19 (11.0) LR-446 (15.1)9 (10.2)37 (17.1)19 (14.5)16 (9.3) LR-5207 (68.1)63 (71.6)144 (66.7)86 (65.6)123 (71.5) M31 (10.2)14 (15.9)17 (7.9)14 (10.7)14 (8.1)AP Arterial phase, APHE Arterial phase hyperenhancement, HBP Hepatobiliary phase, MTM + HCC Macrotrabecular-massive hepatocellular carcinomap-value < 0.05 indicates a significant difference between MTM + HCC and MTM- HCC in the training seta p-value < 0.05 indicates a significant difference between the training and external test setb p-value < 0.05 indicates a significant difference between the internal validation and external test set Radiologic characteristics in the training, internal validation, and external test sets AP Arterial phase, APHE Arterial phase hyperenhancement, HBP Hepatobiliary phase, MTM + HCC Macrotrabecular-massive hepatocellular carcinoma p-value < 0.05 indicates a significant difference between MTM + HCC and MTM- HCC in the training set a p-value < 0.05 indicates a significant difference between the training and external test set b p-value < 0.05 indicates a significant difference between the internal validation and external test set After univariable analysis, infiltrative appearance (p = 0.001), blood product in mass (p < 0.001), fat in mass (p = 0.001), washout pattern (p = 0.017), type of capsule (p = 0.048), intratumoural artery (p < 0.001), substantial necrosis (p < 0.001), central scar (p = 0.024), peritumoral hypointensity in the HBP (p < 0.001), and LI-RADS category (p = 0.022) were found to be associated with MTM + HCC. Segmentation performance of nnU-Net To evaluate segmentation accuracy, we compared the nnU-Net-generated masks (n = 254) with the ground truth segmentations verified by the senior radiologist (Y.Y., with 12 years of experience). In addition to the Dice similarity coefficient, we calculated the 95th percentile Hausdorff distance to assess boundary precision. Overall, Dice similarity coefficients were 0.891 for AP, 0.883 for PP, and 0.932 for HBP, while 95th percentile Hausdorff distance values were 13.4 mm for AP, 13.2 mm for PP, and 10.4 mm for HBP. To further evaluate robustness across different tumor sizes, we stratified the performance by lesion size (≤ 3 cm, 3–5 cm, and > 5 cm). The results are summarized in Supplementary Table S2. Performance of models Based on multivariable logistic regression, ALT (odds ratio [OR] = 2.435 [95% CI: 1.242–4.772], p = 0.010), AFP (OR = 2.839 [95% CI: 1.584–5.085], p < 0.001), maximum tumor diameter (OR = 1.782 [95% CI: 1.329–2.389], p < 0.001), fat in mass (OR = 0.159 [95% CI: 0.065–0.389], p < 0.001), type of capsule (enhancing capsule: OR = 1.782 [95% CI: 0.944–3.366], p = 0.075; nonenhancing capsule: OR = 2.343 [95% CI: 0.953–5.759], p = 0.064) (no capsule as reference) and peritumoral hypointensity in the HBP (OR = 2.675 [95% CI: 1.234–5.799], p = 0.013) were selected for CR model construction (Supplementary Table S3). The CR model had AUCs of 0.801 (95% CI: 0.746–0.856) in the training set, 0.651 (95% CI: 0.563–0.732) in the internal validation set, and 0.739 (95% CI: 0.666–0.803) in the external test set. Finally, we established four habitat radiomics models and four DL models based on AP, PP, HBP, and CP features. Among all habitat radiomics models, the CP model achieved the best performance, with AUCs of 0.802 (95% CI: 0.747–0.858) in the training set, 0.687 (95% CI: 0.585–0.789) in the internal validation set, and 0.681 (95% CI: 0.553–0.808) in the external test set. The CP DL model demonstrated the highest AUC among all DL models, achieving AUCs of 0.740 (95% CI: 0.676–0.804), 0.712 (95% CI: 0.618–0.805), and 0.725 (95% CI: 0.624–0.825) in the training, internal validation, and external test sets, respectively. All habitat radiomics models and DL models were evaluated for diagnostic performance and comparison. The results are shown in Supplementary Appendix S4, Tables S4–S6, and Fig. S4. The correlations between the features of two CP models are presented in Supplementary Fig. S5. In the training, internal validation, and external test sets, median DL scores for the MTM + HCC group were 0.569, 0.478, and 0.518, respectively—significantly higher than those of the MTM − HCC group (0.377, 0.415, and 0.369; p < 0.001 for all sets). Similarly, median habitat radiomics scores for MTM + HCC were 0.591, 0.492, and 0.373 in the training, internal validation, and external test sets, versus 0.334, 0.362, and 0.266 for MTM − HCC (training set: p < 0.001; internal validation set: p = 0.001; external test set: p = 0.002). Comparisons of DL and habitat radiomics scores between MTM+ and MTM − HCC groups are presented in Supplementary Table S7. In the CAM of the Swin Transformer (Fig. 3), the red areas highlighted the regions that the model prioritized most during the prediction task. a AP, (b) PP, and (c) HBP MRI images showing a tumor located in the left liver lobe of a 54-year-old male patient with macrotrabecular-massive hepatocellular carcinoma (MTM + HCC). d Volume of Interest delineated during the AP. Different colors in the figure represent different sub-regions. The red area corresponds to the markedly enhanced portion of the tumor, while the blue area indicates a small amount of necrosis within the tumor. The green area is distributed within the tumor dense cellular areas. e Hematoxylin-eosin staining (original magnification, ×100) of MTM + HCC reveals the macro-trabecular massive pattern. f The Weighted Class Activation Mapping heatmaps for the Swin Transformer model in AP. The red regions indicate areas of activation associated with MTM + HCC Performance and Interpretability of the nomogram After multivariable analysis, habitat radiomics score (OR = 3.561 [95% CI: 2.420–5.241], p < 0.001), DL score (OR = 3.060 [95% CI: 2.092–4.474], p < 0.001), AFP (OR = 3.569 [95% CI: 1.779–7.160], p < 0.001), ALT (OR = 2.419 [95% CI: 1.097–5.336], p = 0.029), and fat in mass (OR = 0.149 [95% CI: 0.053–0.414], p < 0.001) were independent predictors of MTM + HCC (Supplementary Table S8 and Fig. 4a). The correlation between the features of the DL habitat radiomics fusion nomogram was presented in Supplementary Fig. S6. a Forest plot for identifying MTM + HCC in the DL habitat radiomics fusion nomogram. b The DL habitat radiomics fusion nomogram for identifying MTM + HCC preoperatively. AFP, Alpha-fetoprotein; ALT, Alanine aminotransferase; DL, Deep learning; Hb-Rad, Habitat radiomics; MTM, Macrotrabecular-massive The nomogram (Fig. 4b) showed excellent performance with AUCs of 0.897 (95% CI: 0.854–0.940), 0.757 (95% CI: 0.674–0.827), and 0.858 (95% CI: 0.797–0.906) in the training, internal validation, and external test sets, respectively (Table 3). Table 3Comparison of CR model and DL habitat radiomics fusion nomogram for predicting MTM + HCCVariablesAUC [95% CI]Brier scoreECECalibration slopeF1 score [95% CI]sensitivity [95% CI]specificity [95%CI]Accuracyz-valuep-valueTraining set CR model0.801 [0.746–0.856]0.1540.0471.8850.619 [0.535–0.696]0.636 [0.532–0.734]0.829 [0.776–0.878]0.7734.211< 0.0001 DL habitat radiomics fusion nomogram0.897 [0.854–0.940]0.1050.0450.9890.768 [0.698–0.830]0.807 [0.718–0.888]0.880 [0.834–0.919]0.859Internal validation set CR model0.651 [0.563–0.732]0.2000.1280.5360.504 [0.383–0.606]0.882 [0.764–0.977]0.433 [0.333–0.528]0.5502.3830.0172 DL habitat radiomics fusion nomogram0.757 [0.674–0.827]0.1770.1000.4130.566 [0.449–0.667]0.882 [0.769–0.974]0.567 [0.468–0.663]0.649External test set CR model0.739 [0.666–0.803]0.1270.0761.5430.425 [0.281–0.548]0.586 [0.393–0.769]0.762 [0.692–0.827]0.7322.4300.0151 DL habitat radiomics fusion nomogram0.858 [0.797–0.906]0.0930.0591.0370.550 [0.413–0.667]0.862 [0.684–0.960]0.741 [0.683–0.821]0.762AUC Area under the receiver operating characteristic curve, CI Confidence interval, CR Clinical-radiologic model, DL Deep learning, ECE Expected calibration error, MTM + HCC Macrotrabecular-massive hepatocellular carcinoma Comparison of CR model and DL habitat radiomics fusion nomogram for predicting MTM + HCC AUC Area under the receiver operating characteristic curve, CI Confidence interval, CR Clinical-radiologic model, DL Deep learning, ECE Expected calibration error, MTM + HCC Macrotrabecular-massive hepatocellular carcinoma As shown in Fig. 5a, the calibration curve indicated that the nomogram had good diagnostic ability across three datasets. Decision curve analysis showed that the nomogram had good clinical utility in identifying MTM + HCC (Fig. 5b). To further analyze the performance of the nomogram, we constructed a confusion matrix based on the predictions (Supplementary Fig. S7). Calibration curves (a), decision curve analysis (b), and receiver operating characteristic curves (c) of the DL habitat radiomics fusion nomogram and the CR model in the training, internal validation, and external test sets. Recurrence-free survival curves (d) for the predicted MTM+ and predicted MTM- patients identified by the DL habitat radiomics fusion nomogram in the training, internal validation, and external test sets. AUC, Area under the receiver operating characteristic curve; DL, Deep learning; MTM, Macrotrabecular-massive There is a statistically significant improvement of the performance between the AUC of the nomogram and the CR model in the training (0.801 [95% CI: 0.746–0.856] versus 0.897 [95% CI: 0.854–0.940], p < 0.001), internal validation (0.651 [95% CI: 0.563–0.732] versus 0.757 [95% CI: 0.674–0.827], p = 0.017), and external test sets (0.739 [95% CI: 0.666–0.803] versus 0.858 [95% CI: 0.797–0.906], p = 0.015) (Table 3 and Fig. 5c). SHAP analysis (Supplementary Fig. S8) revealed positive contributions from habitat radiomics score, DL score, AFP, and ALT, and a negative contribution from fat mass to the prediction of MTM + HCC. Cox regression analysis for RFS As of December 30, 2024, 536 of 607 patients (88.3%) were included for RFS analysis, with 71 patients lost to follow-up. Across all centers, recurrence occurred in 231 cases (43.1%). Patients with p&&redicted MTM + HCC demonstrated significantly poorer RFS (training set: 18.9 months (95% CI: 15.8–22.0), internal validation set: 26.3 months (95% CI: 20.3–32.3), external test set: 30.0 months (95% CI: 22.8–37.1) compared to those with predicted MTM- HCC (training set: 39.3 months (95% CI: 31.7–46.9), internal validation set: 40.5 months (95% CI: 27.2–53.9), external test set: 46.4 months (95% CI: 36.2–56.6)) across all sets (training set: p < 0.001, internal validation set: p = 0.022, external test set: p = 0.033). Kaplan–Meier curves of RFS are shown in Fig. 5d. As detailed in Supplementary Table S9, the type and method of surgery were comparable across all study groups, with no significant differences observed in any of the three comparisons (MTM + HCC versus MTM- HCC: p = 0.546, 0.152; training set versus external test set: p = 0.131, 0.253; internal validation set versus external test set: p = 0.796, 0.338). In univariable analysis, AFP, ALT, AST, diameter, GGT and nomogram-predicted state were associated with RFS (Table 4). After multivariable analysis, AFP (HR = 2.76, 95% CI: 1.74–4.36, p < 0.001) and nomogram-predicted MTM state (HR = 1.72, 95% CI: 1.03–2.88, p = 0.039) were identified as independent prognostic factors and used to develop the model for RFS. The model achieved C-indexes of 0.74 (95% CI: 0.67–0.81), 0.62 (95% CI: 0.49–0.76), and 0.63 (95% CI: 0.55–0.71) in the training, internal validation, and external test sets. Table 4Univariable and multivariable Cox regression analysis for RFS in the training setCharacteristicsUnivariable analysisMultivariable analysisHR95% CIp-valueHR95% CIp-valueAFP3.842.54–5.82< 0.0012.761.74–4.36< 0.001ALT1.721.13–2.630.0121.060.63–1.790.83AST1.971.32–2.92< 0.0011.050.64–1.710.849Cirrhosis0.820.55–1.210.313Diameter1.131.07–1.19< 0.0011.030.96–1.100.375GGT2.191.49–3.23< 0.0011.390.87–2.210.174Viral hepatitis Noref HBV0.760.49–1.180.225 HCV1.450.44–4.780.541 HBV + HCV co-infection1.060.14–7.870.951 Number1.240.81–1.900.317 Nomogram-predicted MTM state3.222.16–4.81< 0.0011.721.03–2.880.039 Sex1.560.92–2.630.096 Type of surgery Right hepatectomyref Left hepatectomy0.770.45–1.290.316 Extended hepatectomy1.420.80–2.520.230 Right or left lobectomy0.960.55–1.670.887 Right or left segmentectomy0.470.20–1.070.073Method of surgery Open surgeryref Laparoscopic surgery1.450.97–2.160.069Nomogram-predicted MTM state including predicted MTM + HCC and predicted MTM- HCC. AFP Alpha-fetoprotein, ALT Alanine aminotransferase, AST Aspartate aminotransferase, CI confidence interval, GGT Gamma-glutamyl transferase, HR Hazard ratio, MTM + HCC Macrotrabecular-massive hepatocellular carcinoma Univariable and multivariable Cox regression analysis for RFS in the training set Nomogram-predicted MTM state including predicted MTM + HCC and predicted MTM- HCC. AFP Alpha-fetoprotein, ALT Alanine aminotransferase, AST Aspartate aminotransferase, CI confidence interval, GGT Gamma-glutamyl transferase, HR Hazard ratio, MTM + HCC Macrotrabecular-massive hepatocellular carcinoma Discussion MTM + HCC, characterized by its highly aggressive biological behavior and poor prognosis [24], poses significant challenges in clinical strategies. In this study, we developed the DL habitat radiomics fusion nomogram based on Gd-EOB-DTPA-enhanced MRI, which demonstrated superior diagnostic performance compared to the CR model. Notably, patients stratified as MTM + HCC by the nomogram exhibited significantly shorter RFS than those predicted as MTM- HCC. Our study confirmed that AFP is an independent predictor for identifying MTM + HCC, which is consistent with previous findings [25], indicating the high aggressiveness of MTM + HCC. Cheng et al [26] found that the presence of fat in the tumor was an independent predictor for identifying MTM + HCC, which was consistent with our study. Previous studies have demonstrated that fat in mass is more frequently observed in less aggressive HCC and is generally associated with a better prognosis [27–29]. As a highly aggressive subtype, MTM + HCC typically lacks this characteristic. Our study also revealed that elevated ALT levels serve as an independent predictor for identifying MTM + HCC. Previous studies demonstrated that elevated ALT levels are associated with adverse outcomes in HCC patients [30–32]. Elevated ALT levels, indicative of severe hepatic dysfunction, were associated with the aggressive nature of MTM + HCC. Li et al [20] established a DL radiomics model based on dual-energy CT, achieving excellent diagnostic performance with an AUC of 0.89 in the external test set. Our study built various models based on Gd-EOB-DTPA-enhanced MRI, which is widely used in HCC diagnosis. In our study, the performance of the DL habitat radiomics fusion nomogram was superior to that of the CR model. Li et al [20] also found that the DL radiomics model showed better performance than the CR model. The CR model, integrating clinical and qualitative imaging features, was inherently constrained by subjective interpretation and limited variables. In recent years, radiomics methods have shown high efficacy in identifying MTM + HCC. Zhang et al [33] constructed a clinical-radiomics model for MTM + HCC with an AUC of 0.805, but their single-center design limited generalizability. Zhang et al [34] developed a multiphasic CT model with an AUC of 0.836, while our DL-based approach better captures nonlinear patterns. He et al [35] developed a DL and radiomics model based on contrast-enhanced CT, achieving an AUC of 0.773 in the external test set. In contrast, our study utilized habitat radiomics, which more effectively captured the intratumoral spatial heterogeneity, achieving a higher AUC (0.858, 95% CI: 0.797–0.906) in the external test set. There are some limitations to this study. Firstly, as a retrospective cohort, the study was inherently vulnerable to selection bias. Secondly, given the substantial acquisition heterogeneity across the three centers, no dedicated harmonization method was applied, which may limit the generalizability of our findings. Thirdly, the computational demands of our multi‑level fusion nomogram may hinder real‑time clinical use, and an end‑to‑end DL model was not employed to preserve interpretability. Finally, insufficient survival data precluded assessment of overall survival. Future research should aim to expand the sample size to comprehensively evaluate the predictive value of the nomogram in patients’ survival. In conclusion, our study constructed a DL habitat radiomics fusion nomogram based on Gd-EOB-DTPA-enhanced MRI. The nomogram could serve as a robust non-invasive tool for preoperative MTM + HCC diagnosis. Future work should focus on validating our model in larger, more diverse cohorts and exploring its integration into clinical workflows to improve patient outcomes. Supplementary information Additional file 1: Table S1: The 3.0-T MRI scan sequences and parameters for three medical systems. Table S2: Automatically segmentation performance of nnU-Net compared with ground truth in 254 HCC cases. Table S3: Multivariable analysis for identifying MTM+ HCC in the Clinical-radiologic model. Table S4: Performance of habitat radiomics models for predicting MTM+ HCC. Table S5: Performance of DL models for predicting MTM+ HCC. Table S6: Comparison of habitat radiomics models and DL models across different phases. Table S7: Comparison of habitat radiomics score and DL score between MTM+ and MTM- groups in the training, internal validation, and external test sets. Table S8: Multivariable analysis for identifying MTM+ HCC in the DL habitat radiomics fusion nomogram. Table S9: Type and method of surgery across the training, internal validation, and external test sets. Fig. S1: The loss curves and Dice scores of the nnU‑Net model in the arterial phase (a), portal venous phase (b), and hepatobiliary phase (c). Fig. S2: The Calinski-Harabasz index for different numbers of clusters. Fig. S3: The loss (a) and accuracy (b) curves in the training and validation sets of the 3D Swin Transformer model for arterial phase. Fig. S4: ROC curves of AP, PP, HBP and CP habitat radiomics models in the training (a), internal validation (b), and external test (c) sets. The ROC curves of the AP, PP, HBP and CP deep learning models in the training (d), internal validation (e), and external test (f) sets. ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; AP: Arterial phase; PP: Portal venous phase; HBP: Hepatobiliary phase; CP: Combined phase. Fig. S5:—The heatmaps of the selected features in the combined phase (CP) habitat radiomics model (a) and the CP deep learning model (b). The network plots of the selected features in the CP habitat radiomics model (c) and the CP deep learning model (d). AP: Arterial phase; PP: Portal venous phase; HBP: Hepatobiliary phase. Fig. S6: The heat map of the selected features in the DL habitat radiomics fusion nomogram. AFP: Alpha-fetoprotein; ALT: Alanine aminotransferase; DL: Deep learning. Fig. S7: The confusion matrix for the deep learning habitat fusion radiomics nomogram in the training (a), internal validation (b), and external test (c) sets. Fig. S8: The DL habitat fusion radiomics model interpretability using Shapley Additive Explanation methodology. Summary plots of features values against SHAP values and heatmap plots in the training (a), internal validation (b), and external test (c) sets. DL: Deep learning; ALT: Alanine aminotransferase; AFP: Alpha-fetoprotein. Acknowledgements No generative artificial intelligence tools were used in the preparation of this manuscript. Abbreviations Alpha-fetoprotein Alanine transferase Arterial phase Arterial phase hyperenhancement Aspartate transaminase Area under the (receiver operating characteristic) curve Confidence interval Combined-phase Clinical-radiologic Computed tomography Deep learning Gadolinium-ethoxybenzyl-diethylenetriamine pentaacetic acid γ-glutamyl transferase Hepatobiliary phase Hepatocellular carcinoma Hazard ratio Magnetic resonance imaging Macrotrabecular-massive Portal venous phase Recurrence-free survival Shapley additive explanations Author contributions MZ and WQ contributed equally to data collection, data analysis, and manuscript writing, so they shared the first authorship. TZ and MZ were responsible for experimental design and contributed to manuscript revision. CG and MD participated in the data acquisition. WG participated in the data acquisition and data analysis; HS and CH supervised the research and critically reviewed and edited the manuscript. YY, XW, and YY jointly took responsibility for the overall study design, addressed technical challenges encountered during the research, and provided guidance in manuscript writing. All authors reviewed and approved the final manuscript. Funding This work was supported by the Suzhou Science and Technology Plan Project (SKY2023146) and Suzhou Basic Research Pilot Project (SSD2024083) and China Postdoctoral Science Foundation General Program (2024M752334). Data availability The data used or analyzed during the current study are available from the corresponding author upon reasonable request. Ethics approval and consent to participate This multicenter study was approved by the institutional review boards of all three participating hospitals (Institute 1: the First Affiliated Hospital of Soochow University, No. 2024431; Institute 2: the Nantong Third People’s Hospital, No. EK2023025; Institute 3: the First People’s Hospital of Taicang, No. 2022-ky-203), and informed consent was waived due to its retrospective nature. Consent for publication Consent to publish anonymized images is routinely obtained from patients in our institution in line with our institutional policy. Competing interests The authors declare no conflict of interest. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Mengwei Zhang and Qian Wu contributed equally to this work. Contributor Information Yan Yang, Email: yangyan892025@163.com. Ximing Wang, Email: wangximing1998@163.com. Yixing Yu, Email: yuyixing@163.com. Supplementary information The online version contains supplementary material available at https://doi.org/10.1186/s41747-026-00795-y. References Associated Data Supplementary Materials Additional file 1: Table S1: The 3.0-T MRI scan sequences and parameters for three medical systems. Table S2: Automatically segmentation performance of nnU-Net compared with ground truth in 254 HCC cases. Table S3: Multivariable analysis for identifying MTM+ HCC in the Clinical-radiologic model. Table S4: Performance of habitat radiomics models for predicting MTM+ HCC. Table S5: Performance of DL models for predicting MTM+ HCC. Table S6: Comparison of habitat radiomics models and DL models across different phases. Table S7: Comparison of habitat radiomics score and DL score between MTM+ and MTM- groups in the training, internal validation, and external test sets. Table S8: Multivariable analysis for identifying MTM+ HCC in the DL habitat radiomics fusion nomogram. Table S9: Type and method of surgery across the training, internal validation, and external test sets. Fig. S1: The loss curves and Dice scores of the nnU‑Net model in the arterial phase (a), portal venous phase (b), and hepatobiliary phase (c). Fig. S2: The Calinski-Harabasz index for different numbers of clusters. Fig. S3: The loss (a) and accuracy (b) curves in the training and validation sets of the 3D Swin Transformer model for arterial phase. Fig. S4: ROC curves of AP, PP, HBP and CP habitat radiomics models in the training (a), internal validation (b), and external test (c) sets. The ROC curves of the AP, PP, HBP and CP deep learning models in the training (d), internal validation (e), and external test (f) sets. ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; AP: Arterial phase; PP: Portal venous phase; HBP: Hepatobiliary phase; CP: Combined phase. Fig. S5:—The heatmaps of the selected features in the combined phase (CP) habitat radiomics model (a) and the CP deep learning model (b). The network plots of the selected features in the CP habitat radiomics model (c) and the CP deep learning model (d). AP: Arterial phase; PP: Portal venous phase; HBP: Hepatobiliary phase. Fig. S6: The heat map of the selected features in the DL habitat radiomics fusion nomogram. AFP: Alpha-fetoprotein; ALT: Alanine aminotransferase; DL: Deep learning. Fig. S7: The confusion matrix for the deep learning habitat fusion radiomics nomogram in the training (a), internal validation (b), and external test (c) sets. Fig. S8: The DL habitat fusion radiomics model interpretability using Shapley Additive Explanation methodology. Summary plots of features values against SHAP values and heatmap plots in the training (a), internal validation (b), and external test (c) sets. DL: Deep learning; ALT: Alanine aminotransferase; AFP: Alpha-fetoprotein. Data Availability Statement The data used or analyzed during the current study are available from the corresponding author upon reasonable request.