Transformer-Based Deep Learning for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma

Transformer-Based Deep Learning for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Transformer-Based Deep Learning for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma Ruilin He, Huilin Chen, Wenjie Zhou, Mengting Gu, Xingyu Zhao, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7131534/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Microvascular invasion (MVI) is a critical prognostic factor in hepatocellular carcinoma (HCC), but preoperative three-class prediction remains challenging. Radiomics and clinical biomarkers may enable more accurate and individualized assessment. Aim To develop and validate a Transformer-based deep learning framework that integrates radiomic and clinical features for direct three-class MVI classification in HCC patients. Methods This retrospective study included 438 patients with pathologically confirmed HCC and MVI status from a single institution. Radiomic features were extracted from preoperative Gd-BOPTA-enhanced MRI, and clinical laboratory data were collected. A two-stage feature selection strategy, combining univariate statistical testing and recursive feature elimination, was applied. A Transformer-based model was built to classify three MVI categories (M0, M1, M2), and its performance was evaluated on internal and external test sets. Results were compared with traditional machine learning models, including Random Forest, Logistic Regression, XGBoost, and LightGBM. Results The Transformer-based model achieved an accuracy of 0.733, a weighted F1-score of 0.733, and a macro-average AUC of 0.880 (95% CI: 0.807–0.953) on the internal test set. On the external validation set, it reached an accuracy of 0.758, a weighted F1-score of 0.768, and a macro-average AUC of 0.886 (95% CI: 0.833–0.940). It outperformed traditional classifiers and showed superior ability to identify high-risk M2 cases. Conclusions This Transformer-based model enables accurate and objective three-class MVI prediction using multimodal features, supporting individualized surgical planning and improved clinical outcomes. Hepatocellular carcinoma Microvascular invasion Transformer model Radiomics Deep learning Multimodal integration Figures Figure 1 Figure 2 Figure 3 Figure 4 Key Points - A Transformer-based deep learning model was developed for direct three-class MVI prediction in HCC. - The model integrates radiomic features from Gd-BOPTA-enhanced MRI and clinical biomarkers. - It outperforms traditional machine learning models and shows high performance in identifying high-risk M2 cases. Introduction Hepatocellular carcinoma (HCC) is the fifth most common cancer globally and the third leading cause of cancer-related mortality[ 1 ], [ 2 ]. In the Asia–Pacific region, particularly China, chronic hepatitis B virus (HBV) infection accounts for 70–90% of HCC cases [ 3 ]. Although surgical resection and liver transplantation offer potential curative options for early-stage HCC, recurrence rates remain high, affecting 40–70% of patients post-resection and 10–15% post-transplantation, with most recurrences occurring within two years [ 4 ], [ 5 ]. Microvascular invasion (MVI), defined as the presence of tumor cells within microvessels lined by endothelial cells, is a critical pathological feature associated with early recurrence and poor prognosis after surgery [ 6 ]. MVI is further classified into three categories based on the number and distribution of invaded vessels: M0 (no invasion), M1 (≤ 5 invaded vessels within 1 cm of the tumor margin), and M2 (> 5 invaded vessels or invasion beyond 1 cm) [ 6 ], [ 7 ]. Accurate preoperative identification of MVI, particularly M2, is vital for optimizing surgical margins and guiding adjuvant therapies. Currently, MVI diagnosis relies exclusively on postoperative pathological examination, limiting its utility for preoperative decision-making [ 8 ]. To address this limitation, radiomics- and MRI-based predictive models have emerged as promising non-invasive alternatives. However, most existing studies are restricted to binary classification (MVI presence vs. absence) and lack the capacity to differentiate high-risk M2 status[ 9 ],[ 10 ],[ 11 ]. Radiomics enables the extraction of high-throughput, quantitative features from medical images, offering insights into tumor heterogeneity that are imperceptible to the human eye[ 12 ]. When combined with clinical laboratory indicators, radiomics can offer a comprehensive characterization of tumor biology. Nevertheless, the high dimensionality, redundancy, and complex interactions among these features pose challenges for conventional machine learning approaches[ 13 ]. Transformer models, originally developed for natural language processing (NLP) tasks, have recently demonstrated superior performance in modeling high-dimensional structured data through their self-attention mechanisms[ 14 ], [ 15 ]. Their ability to capture non-linear and hierarchical feature dependencies makes them particularly suitable for integrating multimodal data sources, such as radiomic features and clinical variables, which traditional models often struggle with. In this study, we propose a Transformer-based deep learning framework that integrates radiomic features from preoperative MRI and clinical laboratory data for direct three-class MVI classification (M0, M1, M2) in HCC patients. Our approach addresses the limitations of previous studies by enabling fine-grained risk stratification in a fully automated, objective manner. We hypothesize that this method can provide a valuable, non-invasive tool to support individualized surgical planning and improve long-term survival outcomes. While previous radiomics-based approaches primarily focused on binary classification, few have explored direct three-class MVI prediction using deep learning. Moreover, most prior methods relied on conventional classifiers such as logistic regression or tree-based models. In this study, we implemented a Transformer-based architecture specifically designed to handle structured radiomic and clinical features. The model leverages self-attention mechanisms to learn feature interactions and dependencies, thereby enhancing predictive capacity for fine-grained MVI stratification. Materials and methods 1. Patient Enrollment and Clinical Data Collection This retrospective study was approved by the Ethics Committee of Eastern Hepatobiliary Surgery Hospital, with approval number EHBHKY2022-H-P002. The requirement for written informed consent was waived due to its retrospective nature. Patient privacy was strictly protected through data anonymization procedures. A total of 653 patients with hepatocellular carcinoma (HCC) and available microvascular invasion (MVI) analysis were initially identified via electronic medical records using targeted keywords such as “liver,” “tumor,” “mass,” “surgical records,” and “pathological diagnosis.” Among them, 458 patients were from Hospital A and 195 patients from Hospital B. After applying predefined inclusion and exclusion criteria, 438 eligible patients remained, including 305 from Hospital A and 133 from Hospital B. From Hospital A, 229 patients were randomly allocated to the training set, and 76 patients to the internal test set. The remaining 133 patients from Hospital B were assigned to an independent external validation set to evaluate the generalizability of the model across institutions. The inclusion criteria were as follows: Patients who underwent liver resection and received Gd-BOPTA-enhanced MRI within two months prior to surgery. Histopathological confirmation of hepatocellular carcinoma (HCC) and microvascular invasion (MVI) status. No prior treatments before surgery, such as radiofrequency ablation or transarterial chemoembolization (TACE). Absence of extrahepatic malignancies. Availability of complete clinical, imaging, and pathological data. Exclusion criteria included patients who underwent preoperative local therapies, had extrahepatic malignancies, or lacked essential clinical or imaging data. A detailed patient selection workflow is presented in Fig. 1 . 2. MRI Image Acquisition All participants were required to fast for at least 6 hours and to abstain from water intake for at least 4 hours before the MRI examination. The scanning range extended from the upper edge to the lower edge of the liver. The contrast agent, 0.1 mmol/kg of Gd-BOPTA (MultiHance, Bracco), was administered through injection into the median vein of the patient's elbow using a high-pressure syringe at a rate of 2.0 mL/s, followed by a 20-mL saline flush lasting approximately 20 seconds. After contrast agent administration, imaging was performed during specific phases: arterial phase (AP) at 25–45 seconds, portal venous phase (PVP) at 50–70 seconds, and delayed phase (DP) at 100–180 seconds. The MRI protocol also included fat-suppressed T2-weighted imaging (T2WI), T1-weighted imaging (T1WI) and diffusion-weighted imaging (DWI). 3. Imaging Analysis Manual segmentation of HCC lesions was conducted to construct the radiomics dataset. Two radiologists, each with more than five years of experience in abdominal MRI interpretation, independently delineated the region of interest (ROI) of HCC tumors on MRI. Subsequently, a third senior radiologist with over ten years of experience reviewed and verified these ROIs to ensure accuracy and consistency. Manual segmentation was performed using ITK-SNAP (version 3.6.0; http://www.itksnap.org ) software. The delineations were based on the visible tumor boundaries in MRI sequences, including T2WI, T1WI, DWI, AP, PVP, DP. The final consensus ROIs were subsequently used for radiomics feature extraction and statistical analysis. Importantly, no qualitative assessments assessed by radiologists were included in this study. All data inputs for model development were based solely on objective information, including clinical laboratory parameters and automatically extracted radiomics features from manually delineated ROIs. This approach minimizes potential interobserver variability and enhances the objectivity, the reproducibility, and the generalizability of the predictive model. 4. Pathological evaluation All hematoxylin and eosin (H&E)-stained sections were retrospectively reviewed for MVI by two board-certified pathologists, each with more than 10 years of experience in hepatopathology. The pathologists were blinded to all clinical and imaging data during the assessment. In cases of disagreement, a senior pathologist with over 20 years of diagnostic experience was consulted to reach a consensus. MVI status was determined according to standardized pathological criteria. Patients were classified into three groups: M0 (no MVI), M1 (1–5 MVI sites within ≤ 1 cm from the tumor margin), and M2 (> 5 MVI sites or any MVI detected beyond 1 cm from the tumor margin). This grading system has been widely adopted in clinical research for prognostic stratification of HCC[ 16 ]. 5. Radiomic Feature Extraction Radiomic feature extraction was conducted following manual segmentation of the region of interest (ROI) by an experienced radiologist using ITK-SNAP software (version 3.6.0)[ 17 ], [ 18 ]. The delineated ROIs were used to extract quantitative features with the PyRadiomics library. A total of 1,500 radiomic features were obtained, including first-order statistical descriptors (such as mean, skewness, and kurtosis), shape-based metrics (such as sphericity and elongation), and texture features derived from gray-level matrices, including the Gray-Level Co-occurrence Matrix (GLCM) and the Gray-Level Run-Length Matrix (GLRLM)[ 18 ]. To better characterize tumor heterogeneity, features were extracted not only from the original images but also from filtered versions processed via wavelet and Laplacian of Gaussian (LoG) transformations[ 19 ], [ 20 ]. 6. Data Preprocessing Preprocessing involved two major steps: Missing Value Handling : Samples with more than 10% missing values were excluded. For the remaining samples, missing values were imputed separately for each feature using k-nearest neighbor (KNN) imputation (k = 3), based on the Euclidean distance in the feature space[ 21 ], [ 22 ]. Normalization : Radiomic features were normalized using Z-score transformation. The mean and standard deviation were calculated exclusively from the training set and then applied to the validation and test sets to ensure no information leakage during model evaluation[ 23 ]. 7. Feature Selection To mitigate the risk of overfitting and enhance model generalizability, a two-stage feature selection strategy was employed. In the initial screening phase, features were first evaluated for reproducibility using the intraclass correlation coefficient (ICC)[ 19 ], and only those with ICC > 0.75 were retained. Subsequently, univariate statistical tests (t-test or Mann-Whitney U test, depending on data distribution) were performed to exclude features with weak discriminative ability. In the dimensionality reduction phase, Recursive Feature Elimination (RFE)[ 24 ] based on Random Forest classifiers was conducted to further select the most informative subset of features. To systematically evaluate the impact of different feature selection strategies, sixteen pipelines (4 initial screening methods × 4 dimensionality reduction methods) were constructed and compared using cross-validation performance within the training set. Ultimately, the pipeline combining “t-test or Mann-Whitney U test + RFE with Random Forest” was selected for final model training, as it achieved the best balance between feature compactness and model performance. 8. Model Architecture and Training We developed a Transformer-based deep learning model tailored for three-class MVI prediction. The model architecture was adapted from the Transformer framework, originally introduced for natural language processing tasks[ 14 ], and later extended to structured tabular data such as clinical and radiomic features[ 15 ]. As illustrated in Fig. 3 , the model architecture comprised three main components: an input embedding layer that projected the input features into an 8-dimensional latent space; two stacked Transformer encoder layers, each utilizing 2 attention heads, a feedforward dimension of 16, and a dropout rate of 0.2; and a final output classification layer that generated predictions across three categories. Following the definition of the model architecture, we proceeded to model training. Cross-entropy loss with class-balanced weights was employed to mitigate class imbalance. Model parameters were optimized using the AdamW optimizer with a learning rate of 0.0005 and a weight decay of 0.02. The mini-batch size was set to 16. Training was conducted for 2000 epochs on an NVIDIA RTX 4090 GPU. 9. Data Augmentation and Balancing Based on the extracted radiomic features, the overall preprocessing and modeling strategy is summarized in Fig. 3 . This includes feature normalization, missing value imputation, two-step feature selection, and Transformer-based classification. To mitigate the issue of class imbalance in the training set, a hybrid resampling strategy was employed. Specifically, the Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN) was utilized to simultaneously oversample minority classes and clean borderline or noisy samples[ 22 ]. Following resampling, updated class distributions were computed, and class weights were recalculated and incorporated into the cross-entropy loss function to further mitigate residual imbalance during model training. This combined approach has been demonstrated to enhance classifier performance in highly imbalanced multiclass medical datasets. Results Model Performance Across Datasets The proposed Transformer-based model demonstrated robust classification performance across training, test, and validation sets. On the training set, the model achieved a loss of 0.540, an accuracy of 0.759, a weighted F1-score of 0.766, and a macro-average AUC of 0.920. On the independent test set, the model yielded a loss of 0.816, an accuracy of 0.733, a weighted F1-score of 0.733, and a macro-average AUC of 0.880 (95% CI: 0.807–0.953). On the external validation set, it reached a loss of 0.948, an accuracy of 0.758, a weighted F1-score of 0.768, and a macro-average AUC of 0.886 (95% CI: 0.833–0.940). The confusion matrices and corresponding ROC curves for all three datasets are presented in Fig. 4 . The model consistently demonstrated strong discriminative ability for Class 0 across all sets, while Class 1 and Class 2 showed moderate variation in precision and recall. Summary of Performance Metrics A comprehensive summary of model performance across the training, validation, and test sets is provided in Table 1 . Table 1 Overall performance of the proposed Transformer-based model. Dataset Accuracy Weighted F1-score Macro-average Precision Macro-average Recall Macro-average AUC (95% CI) Training Set 0.759 0.766 0.761 0.794 0.920 (0.884–0.955) Validation Set 0.758 0.769 0.676 0.719 0.886 (0.833–0.940) Test Set 0.733 0.733 0.635 0.635 0.880 (0.807–0.953) Ablation Study with Traditional Machine Learning Models To assess the effectiveness of the Transformer-based model, we compared its performance against several traditional machine learning classifiers, including Random Forest, Logistic Regression, XGBoost, and LightGBM. As summarized in Table 2 , the Transformer-based model achieved the highest accuracy, F1-score, and AUC on both the validation and test sets, outperforming all traditional models across all evaluation metrics. Table 2. Comparison of the proposed Transformer-based model with traditional machine learning models. Model Validation Accuracy Validation F1-score Validation AUC Test Accuracy Test F1-score Test AUC Random Forest 0.682 0.694 0.834 0.640 0.660 0.827 Logistic Regression 0.629 0.638 0.837 0.547 0.575 0.802 XGBoost 0.545 0.568 0.807 0.533 0.557 0.804 LightGBM 0.523 0.547 0.800 0.573 0.596 0.823 Transformer (ours) 0.758 0.769 0.886 0.733 0.733 0.880 Discussion Our adoption of a Transformer-based model in this context constitutes a methodological innovation in MVI classification. While Transformer architecture has been increasingly applied in medical imaging, their application to structured radiomic data for multi-class classification remains underrepresented. By tailoring the model to capture dependencies among quantitative features, our framework extends beyond prior binary MVI classification attempts, offering a more nuanced and clinically useful risk stratification strategy. This study demonstrated the feasibility and effectiveness of using a Transformer-based deep learning model to perform direct three-class classification of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) patients by integrating radiomic features and clinical laboratory indicators. Compared with conventional machine learning methods such as Random Forest and Logistic Regression, the proposed model consistently achieved superior performance, with a test set AUC of 0.880 (95% CI: 0.807–0.953) and a weighted F1-score of 0.733. These findings highlight the Transformer architecture’s capacity to model complex, high-dimensional, and structured medical data by capturing nonlinear and hierarchical feature dependencies through its self-attention mechanism[ 14 ], [ 15 ], [ 25 ]. Previous radiomics-based studies have primarily focused on binary MVI classification (presence vs. absence), typically reporting AUCs ranging from 0.75 to 0.85[ 10 ], [ 26 ], [ 27 ]. In contrast, our approach enables one-step, three-class prediction (M0, M1, M2), providing a more granular and clinically actionable preoperative risk stratification strategy. This advancement allows for refined surgical planning and supports individualized treatment decisions. Additionally, unlike earlier studies that often relied on subjective imaging features manually assessed by radiologists, our model is entirely based on objectively extracted features, which reduces interobserver variability and improves reproducibility[ 28 ]. The key contributions of this study lie in three aspects: proposing a Transformer-based framework for direct MVI three-class classification, demonstrating the benefit of multimodal feature integration from radiomics and clinical data, and ensuring high model reproducibility by using only objective quantitative inputs. Nevertheless, several limitations must be acknowledged. First, although the model performed well on both internal and external validation sets, the dataset was collected from a limited geographic region, which may restrict its generalizability to broader populations and imaging protocols[ 29 ]. Second, the proportion of M2 cases remained low (~ 15%), and although SMOTE-ENN[ 22 ] and class-weighted loss were applied, the class imbalance still posed challenges in reliably distinguishing M1 from M2. Third, as with most deep learning models, the Transformer lacks interpretability, which may hinder its acceptance in clinical decision-making[ 30 ]. Future research should focus on external validation using multi-center and multi-ethnic datasets to improve robustness and generalizability[ 31 ]. In addition, incorporating multi-omics data (e.g., genomics, transcriptomics, proteomics) may further enhance the model’s predictive capacity and biological interpretability[ 32 ]. To address the black-box nature of the Transformer, explainable artificial intelligence (XAI) techniques—such as SHAP value analysis, attention heatmaps, and counterfactual reasoning—should be integrated[ 33 ], [ 34 ]. Exploring advanced Transformer variants, including sparse attention mechanisms or hybrid CNN-Transformer architectures, may also contribute to improved efficiency and clinical utility[ 35 ]. Conclusion In this study, we proposed a Transformer-based deep learning framework that integrates radiomic features and clinical laboratory indicators for the preoperative three-class classification of MVI in patients with HCC. The proposed model effectively captured complex high-dimensional feature relationships and achieved competitive performance across training, validation, and independent test sets. By enabling direct three-class MVI prediction in a single step, our approach provides more granular prognostic information and holds promise for supporting individualized treatment planning and streamlining clinical decision-making. Despite the encouraging results, further work is needed to enhance the model’s clinical applicability. Future research should focus on expanding validation cohorts to larger, multi-center, and multi-ethnic datasets, improving model interpretability through explainable artificial intelligence (XAI) techniques, and incorporating multi-omics data to enrich biological insights and further boost predictive accuracy. Abbreviations HCC Hepatocellular carcinoma MVI Microvascular invasion MRI Magnetic resonance imaging AUC Area under the curve Gd-BOPTA Gadolinium benzyloxypropionictetraacetate RFE Recursive feature elimination Declarations Author Contribution JN and WL conceived and designed the study. CH was responsible for data collection, and contributed to ROI segmentation and radiomics processing along with ZW, GM, and ZX. HR developed the model, performed the data analysis, and drafted the manuscript. JN and WL provided supervision and revised the manuscript critically for important intellectual content. All authors read and approved the final version of the manuscript. Data Availability The data used in this study were obtained from the Eastern Hepatobiliary Surgery Hospital and are not publicly available due to institutional and ethical restrictions. Access to the data can be requested from the corresponding author and may be granted upon reasonable request and approval from the institutional review board. Critical Relevance Statement Accurate preoperative three-class prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) remains unmet in clinical imaging. This study demonstrates the potential of a Transformer-based model for non-invasive and individualized MVI stratification. References H. Sung et al. , “Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries,” CA. Cancer J. Clin. , vol. 71, no. 3, pp. 209–249, May 2021, doi: 10.3322/caac.21660. J. Calderaro, T. P. Seraphin, T. Luedde, and T. G. Simon, “Artificial intelligence for the prevention and clinical management of hepatocellular carcinoma,” J. Hepatol. , vol. 76, no. 6, pp. 1348–1361, Jun. 2022, doi: 10.1016/j.jhep.2022.01.014. Z. Lei et al. , “Nomogram for Preoperative Estimation of Microvascular Invasion Risk in Hepatitis B Virus–Related Hepatocellular Carcinoma Within the Milan Criteria,” JAMA Surg. , vol. 151, no. 4, p. 356, Apr. 2016, doi: 10.1001/jamasurg.2015.4257. B. Kim, J. Kahn, and N. A. Terrault, “Liver transplantation as therapy for hepatocellular carcinoma,” Liver Int. , vol. 40, no. S1, pp. 116–121, Feb. 2020, doi: 10.1111/liv.14346. M. Tampaki, G. V. Papatheodoridis, and E. Cholongitas, “Intrahepatic recurrence of hepatocellular carcinoma after resection: an update,” Clin. J. Gastroenterol. , vol. 14, no. 3, pp. 699–713, Jun. 2021, doi: 10.1007/s12328-021-01394-7. S. Sumie et al. , “The Significance of Classifying Microvascular Invasion in Patients with Hepatocellular Carcinoma,” Ann. Surg. Oncol. , vol. 21, no. 3, pp. 1002–1009, Mar. 2014, doi: 10.1245/s10434-013-3376-9. Y. Cai et al. , “Stathmin 1 is a biomarker for diagnosis of microvascular invasion to predict prognosis of early hepatocellular carcinoma,” Cell Death Dis. , vol. 13, no. 2, p. 176, Feb. 2022, doi: 10.1038/s41419-022-04625-y. M. Rodríguez-Perálvarez, T. V. Luong, L. Andreana, T. Meyer, A. P. Dhillon, and A. K. Burroughs, “A Systematic Review of Microvascular Invasion in Hepatocellular Carcinoma: Diagnostic and Prognostic Variability,” Ann. Surg. Oncol. , vol. 20, no. 1, pp. 325–339, Jan. 2013, doi: 10.1245/s10434-012-2513-1. W. Jian et al. , “Predicting microvascular invasion in solitary hepatocellular carcinoma: a multi-center study integrating clinical, MRI assessments, and radiomics indicators,” Front. Oncol. , vol. 15, p. 1511260, Feb. 2025, doi: 10.3389/fonc.2025.1511260. G. Hu et al. , “Radiogenomics nomogram based on MRI and microRNAs to predict microvascular invasion of hepatocellular carcinoma,” Front. Oncol. , vol. 14, p. 1371432, Jul. 2024, doi: 10.3389/fonc.2024.1371432. J. Li et al. , “A clinical study exploring the prediction of microvascular invasion in hepatocellular carcinoma through the use of combined enhanced CT and MRI radiomics,” PLOS ONE , vol. 20, no. 1, p. e0318232, Jan. 2025, doi: 10.1371/journal.pone.0318232. P. Lambin et al. , “Radiomics: Extracting more information from medical images using advanced feature analysis,” Eur. J. Cancer , vol. 48, no. 4, pp. 441–446, Mar. 2012, doi: 10.1016/j.ejca.2011.11.036. C. Parmar, P. Grossmann, J. Bussink, P. Lambin, and H. J. W. L. Aerts, “Machine Learning methods for Quantitative Radiomic Biomarkers,” Sci. Rep. , vol. 5, no. 1, p. 13087, Aug. 2015, doi: 10.1038/srep13087. A. Vaswani et al. , “Attention Is All You Need,” Aug. 01, 2023, arXiv : arXiv:1706.03762. Accessed: Nov. 06, 2023. [Online]. Available: http://arxiv.org/abs/1706.03762 X. Huang, A. Khetan, M. Cvitkovic, and Z. Karnin, “TabTransformer: Tabular Data Modeling Using Contextual Embeddings,” Dec. 11, 2020, arXiv : arXiv:2012.06678. doi: 10.48550/arXiv.2012.06678. for the Liver Cancer Pathology Group of China (LCPGC) et al. , “A standardized pathological proposal for evaluating microvascular invasion of hepatocellular carcinoma: a multicenter study by LCPGC,” Hepatol. Int. , vol. 14, no. 6, pp. 1034–1047, Dec. 2020, doi: 10.1007/s12072-020-10111-4. P. A. Yushkevich et al. , “User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability,” NeuroImage , vol. 31, no. 3, pp. 1116–1128, Jul. 2006, doi: 10.1016/j.neuroimage.2006.01.015. J. J. M. Van Griethuysen et al. , “Computational Radiomics System to Decode the Radiographic Phenotype,” Cancer Res. , vol. 77, no. 21, pp. e104–e107, Nov. 2017, doi: 10.1158/0008-5472.CAN-17-0339. A. Zwanenburg et al. , “The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping,” Radiology , vol. 295, no. 2, pp. 328–338, May 2020, doi: 10.1148/radiol.2020191145. C. Parmar et al. , “Robust Radiomics Feature Quantification Using Semiautomatic Volumetric Segmentation,” PLoS ONE , vol. 9, no. 7, p. e102107, Jul. 2014, doi: 10.1371/journal.pone.0102107. O. Troyanskaya et al. , “Missing value estimation methods for DNA microarrays,” Bioinformatics , vol. 17, no. 6, pp. 520–525, Jun. 2001, doi: 10.1093/bioinformatics/17.6.520. G. E. A. P. A. Batista, R. C. Prati, and M. C. Monard, “A study of the behavior of several methods for balancing machine learning training data,” ACM SIGKDD Explor. Newsl. , vol. 6, no. 1, pp. 20–29, Jun. 2004, doi: 10.1145/1007730.1007735. C. Sammut and G. I. Webb, Eds., Encyclopedia of Machine Learning and Data Mining . Boston, MA: Springer US, 2017. doi: 10.1007/978-1-4899-7687-1. I. Guyon, J. Weston, and S. Barnhill, “Gene Selection for Cancer Classification using Support Vector Machines”. G. Somepalli, M. Goldblum, A. Schwarzschild, C. B. Bruss, and T. Goldstein, “SAINT: Improved Neural Networks for Tabular Data via Row Attention and Contrastive Pre-Training,” Jun. 02, 2021, arXiv : arXiv:2106.01342. doi: 10.48550/arXiv.2106.01342. S.-T. Feng et al. , “Preoperative prediction of microvascular invasion in hepatocellular cancer: a radiomics model using Gd-EOB-DTPA-enhanced MRI,” Eur. Radiol. , vol. 29, no. 9, pp. 4648–4659, Sep. 2019, doi: 10.1007/s00330-018-5935-8. M. He, P. Zhang, X. Ma, B. He, C. Fang, and F. Jia, “Radiomic Feature-Based Predictive Model for Microvascular Invasion in Patients With Hepatocellular Carcinoma,” Front. Oncol. , vol. 10, p. 574228, Nov. 2020, doi: 10.3389/fonc.2020.574228. W. Zhang, Y. Guo, and Q. Jin, “Radiomics and Its Feature Selection: A Review,” Symmetry , vol. 15, no. 10, p. 1834, Sep. 2023, doi: 10.3390/sym15101834. R. D. Riley et al. , “External validation of clinical prediction models using big datasets from e-health records or IPD meta-analysis: opportunities and challenges,” BMJ , p. i3140, Jun. 2016, doi: 10.1136/bmj.i3140. A. Holzinger, C. Biemann, C. S. Pattichis, and D. B. Kell, “What do we need to build explainable AI systems for the medical domain?,” Dec. 28, 2017, arXiv : arXiv:1712.09923. doi: 10.48550/arXiv.1712.09923. R. D. Riley et al. , “Evaluation of clinical prediction models (part 2): how to undertake an external validation study,” BMJ , p. e074820, Jan. 2024, doi: 10.1136/bmj-2023-074820. P. Kickingereder et al. , “Radiogenomics of Glioblastoma: Machine Learning–based Classification of Molecular Characteristics by Using Multiparametric and Multiregional MR Imaging Features,” Radiology , vol. 281, no. 3, pp. 907–918, Dec. 2016, doi: 10.1148/radiol.2016161382. S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions”. E. Tjoa and C. Guan, “A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI,” IEEE Trans. Neural Netw. Learn. Syst. , vol. 32, no. 11, pp. 4793–4813, Nov. 2021, doi: 10.1109/TNNLS.2020.3027314. K. Choromanski et al. , “Rethinking Attention with Performers,” Nov. 19, 2022, arXiv : arXiv:2009.14794. doi: 10.48550/arXiv.2009.14794. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7131534","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":490032911,"identity":"1d3183b3-fb0a-4848-947f-56736101b517","order_by":0,"name":"Ruilin He","email":"","orcid":"","institution":"university of shanghai for science and technology","correspondingAuthor":false,"prefix":"","firstName":"Ruilin","middleName":"","lastName":"He","suffix":""},{"id":490032912,"identity":"5e168ba0-9455-44af-a191-229e0f0036ad","order_by":1,"name":"Huilin Chen","email":"","orcid":"","institution":"The Third Affiliated Hospital of Shanghai Naval Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Huilin","middleName":"","lastName":"Chen","suffix":""},{"id":490032913,"identity":"b6f66998-f163-4566-8e83-8a480701f2b4","order_by":2,"name":"Wenjie Zhou","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Wenjie","middleName":"","lastName":"Zhou","suffix":""},{"id":490032914,"identity":"217b0a17-5d56-436a-a623-129c2b4fb46d","order_by":3,"name":"Mengting Gu","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Mengting","middleName":"","lastName":"Gu","suffix":""},{"id":490032915,"identity":"e519dc27-ff96-49ea-af9c-7f8ee84c9d82","order_by":4,"name":"Xingyu Zhao","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Xingyu","middleName":"","lastName":"Zhao","suffix":""},{"id":490032916,"identity":"baea82be-1740-49c2-ab77-ddf69fa5574f","order_by":5,"name":"Ning-Yang Jia","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA2ElEQVRIiWNgGAWjYDACZijN2MzA+CChooY0LcwGD84cI81CNsmHLcyElZmz8x5+zVNzx665nflZRWIDGwN/e3cCXi2WzXxp1jzHniU3NrOZ3UjcIcMgcebsBrxaDA7zmBnzsB1OBvoFqOUMG4OBRC4xWv6BtLB/K0hsYyZKi/Fj3rbDdozNPGYMxGoxY5zbdzgBqKVYIuHMMR7Cfjl/xvjDm2+H7Q37j2/8+KOiRo6/vRe/FiBgk+JhYEjc2ADh8RBSDgLMH38wMNjLE6N0FIyCUTAKRiYAAMfXSFzRDyVJAAAAAElFTkSuQmCC","orcid":"","institution":"university of shanghai for science and technology","correspondingAuthor":true,"prefix":"","firstName":"Ning-Yang","middleName":"","lastName":"Jia","suffix":""},{"id":490032917,"identity":"0d2e8710-bac3-466c-8ca1-9c2af3c5a433","order_by":6,"name":"Wanmin Liu","email":"","orcid":"","institution":"Shanghai Jiao Tong University","correspondingAuthor":false,"prefix":"","firstName":"Wanmin","middleName":"","lastName":"Liu","suffix":""}],"badges":[],"createdAt":"2025-07-15 14:23:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7131534/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7131534/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87728130,"identity":"1044ee09-a773-47c0-9ee1-2af40e89a976","added_by":"auto","created_at":"2025-07-28 11:07:18","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":40700,"visible":true,"origin":"","legend":"\u003cp\u003eThe workflow of patient selection. HCC: hepatocellular carcinoma. MVI: microvascular invasion\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7131534/v1/5a4dc0bfebd638e339b190d2.png"},{"id":87728136,"identity":"ea7226ef-1e4c-4b44-b33a-23d0496a020c","added_by":"auto","created_at":"2025-07-28 11:07:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":109307,"visible":true,"origin":"","legend":"\u003cp\u003eOverall workflow for radiomics-based three-class classification of microvascular invasion (MVI) in hepatocellular carcinoma (HCC). Radiomic features were extracted from Gd-enhanced MRI based on manually delineated regions of interest (ROIs). A two-step feature selection strategy was applied, combining univariate statistical tests (t-test or Mann–Whitney U test) and recursive feature elimination (RFE). A Transformer model was trained for MVI classification (M0, M1, M2), and performance was evaluated using confusion matrices and ROC curves.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7131534/v1/9bc12047124c194d264bbcc0.png"},{"id":87728924,"identity":"abaefa84-82af-4e44-9b96-3a32ae3bf738","added_by":"auto","created_at":"2025-07-28 11:15:18","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":43102,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the Transformer-based model architecture for MVI classification.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7131534/v1/63955b37a244887a807eee16.png"},{"id":87728132,"identity":"e709bb7a-6cc8-414d-89f2-06e18f40bcb8","added_by":"auto","created_at":"2025-07-28 11:07:19","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":83607,"visible":true,"origin":"","legend":"\u003cp\u003eClassification performance of the Transformer-based model across training, test, and validation sets. \u003cstrong\u003eTop row\u003c/strong\u003e: Training set — Confusion matrix (left) and ROC curves (right). \u003cstrong\u003eMiddle row\u003c/strong\u003e: Test set — Confusion matrix (left) and ROC curves (right). \u003cstrong\u003eBottom row\u003c/strong\u003e: External validation set — Confusion matrix (left) and ROC curves (right).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7131534/v1/98c5775073fec4fe8ecf2948.png"},{"id":88331776,"identity":"e5894dd7-bdb9-4bff-8e67-1b2e92e32d0a","added_by":"auto","created_at":"2025-08-05 10:54:11","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1075975,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7131534/v1/0ff6c594-e727-4721-9b43-00f384ec155a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Transformer-Based Deep Learning for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma","fulltext":[{"header":"Key Points","content":"\u003cp\u003e- A Transformer-based deep learning model was developed for direct three-class MVI prediction in HCC.\u003c/p\u003e\u003cp\u003e- The model integrates radiomic features from Gd-BOPTA-enhanced MRI and clinical biomarkers.\u003c/p\u003e\u003cp\u003e- It outperforms traditional machine learning models and shows high performance in identifying high-risk M2 cases.\u003c/p\u003e"},{"header":"Introduction","content":"\u003cp\u003eHepatocellular carcinoma (HCC) is the fifth most common cancer globally and the third leading cause of cancer-related mortality[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. In the Asia\u0026ndash;Pacific region, particularly China, chronic hepatitis B virus (HBV) infection accounts for 70\u0026ndash;90% of HCC cases [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Although surgical resection and liver transplantation offer potential curative options for early-stage HCC, recurrence rates remain high, affecting 40\u0026ndash;70% of patients post-resection and 10\u0026ndash;15% post-transplantation, with most recurrences occurring within two years [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMicrovascular invasion (MVI), defined as the presence of tumor cells within microvessels lined by endothelial cells, is a critical pathological feature associated with early recurrence and poor prognosis after surgery [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. MVI is further classified into three categories based on the number and distribution of invaded vessels: M0 (no invasion), M1 (\u0026le;\u0026thinsp;5 invaded vessels within 1 cm of the tumor margin), and M2 (\u0026gt;\u0026thinsp;5 invaded vessels or invasion beyond 1 cm) [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Accurate preoperative identification of MVI, particularly M2, is vital for optimizing surgical margins and guiding adjuvant therapies.\u003c/p\u003e\u003cp\u003eCurrently, MVI diagnosis relies exclusively on postoperative pathological examination, limiting its utility for preoperative decision-making [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. To address this limitation, radiomics- and MRI-based predictive models have emerged as promising non-invasive alternatives. However, most existing studies are restricted to binary classification (MVI presence vs. absence) and lack the capacity to differentiate high-risk M2 status[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e],[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e],[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eRadiomics enables the extraction of high-throughput, quantitative features from medical images, offering insights into tumor heterogeneity that are imperceptible to the human eye[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. When combined with clinical laboratory indicators, radiomics can offer a comprehensive characterization of tumor biology. Nevertheless, the high dimensionality, redundancy, and complex interactions among these features pose challenges for conventional machine learning approaches[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTransformer models, originally developed for natural language processing (NLP) tasks, have recently demonstrated superior performance in modeling high-dimensional structured data through their self-attention mechanisms[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Their ability to capture non-linear and hierarchical feature dependencies makes them particularly suitable for integrating multimodal data sources, such as radiomic features and clinical variables, which traditional models often struggle with.\u003c/p\u003e\u003cp\u003eIn this study, we propose a Transformer-based deep learning framework that integrates radiomic features from preoperative MRI and clinical laboratory data for direct three-class MVI classification (M0, M1, M2) in HCC patients. Our approach addresses the limitations of previous studies by enabling fine-grained risk stratification in a fully automated, objective manner. We hypothesize that this method can provide a valuable, non-invasive tool to support individualized surgical planning and improve long-term survival outcomes.\u003c/p\u003e\u003cp\u003eWhile previous radiomics-based approaches primarily focused on binary classification, few have explored direct three-class MVI prediction using deep learning. Moreover, most prior methods relied on conventional classifiers such as logistic regression or tree-based models. In this study, we implemented a Transformer-based architecture specifically designed to handle structured radiomic and clinical features. The model leverages self-attention mechanisms to learn feature interactions and dependencies, thereby enhancing predictive capacity for fine-grained MVI stratification.\u003c/p\u003e"},{"header":"Materials and methods","content":"\n\u003ch3\u003e1. Patient Enrollment and Clinical Data Collection\u003c/h3\u003e\n\u003cp\u003e This retrospective study was approved by the Ethics Committee of Eastern Hepatobiliary Surgery Hospital, with approval number EHBHKY2022-H-P002. The requirement for written informed consent was waived due to its retrospective nature. Patient privacy was strictly protected through data anonymization procedures.\u003c/p\u003e\u003cp\u003eA total of 653 patients with hepatocellular carcinoma (HCC) and available microvascular invasion (MVI) analysis were initially identified via electronic medical records using targeted keywords such as \u0026ldquo;liver,\u0026rdquo; \u0026ldquo;tumor,\u0026rdquo; \u0026ldquo;mass,\u0026rdquo; \u0026ldquo;surgical records,\u0026rdquo; and \u0026ldquo;pathological diagnosis.\u0026rdquo; Among them, 458 patients were from Hospital A and 195 patients from Hospital B.\u003c/p\u003e\u003cp\u003e After applying predefined inclusion and exclusion criteria, 438 eligible patients remained, including 305 from Hospital A and 133 from Hospital B.\u003c/p\u003e\u003cp\u003eFrom Hospital A, 229 patients were randomly allocated to the training set, and 76 patients to the internal test set. The remaining 133 patients from Hospital B were assigned to an independent external validation set to evaluate the generalizability of the model across institutions.\u003c/p\u003e\u003cp\u003eThe inclusion criteria were as follows:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003ePatients who underwent liver resection and received Gd-BOPTA-enhanced MRI within two months prior to surgery.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eHistopathological confirmation of hepatocellular carcinoma (HCC) and microvascular invasion (MVI) status.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eNo prior treatments before surgery, such as radiofrequency ablation or transarterial chemoembolization (TACE).\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eAbsence of extrahepatic malignancies.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eAvailability of complete clinical, imaging, and pathological data.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eExclusion criteria included patients who underwent preoperative local therapies, had extrahepatic malignancies, or lacked essential clinical or imaging data.\u003c/p\u003e\u003cp\u003eA detailed patient selection workflow is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003e2. MRI Image Acquisition\u003c/h3\u003e\n\u003cp\u003eAll participants were required to fast for at least 6 hours and to abstain from water intake for at least 4 hours before the MRI examination. The scanning range extended from the upper edge to the lower edge of the liver.\u003c/p\u003e\u003cp\u003eThe contrast agent, 0.1 mmol/kg of Gd-BOPTA (MultiHance, Bracco), was administered through injection into the median vein of the patient's elbow using a high-pressure syringe at a rate of 2.0 mL/s, followed by a 20-mL saline flush lasting approximately 20 seconds.\u003c/p\u003e\u003cp\u003eAfter contrast agent administration, imaging was performed during specific phases: arterial phase (AP) at 25\u0026ndash;45 seconds, portal venous phase (PVP) at 50\u0026ndash;70 seconds, and delayed phase (DP) at 100\u0026ndash;180 seconds. The MRI protocol also included fat-suppressed T2-weighted imaging (T2WI), T1-weighted imaging (T1WI) and diffusion-weighted imaging (DWI).\u003c/p\u003e\n\u003ch3\u003e3. Imaging Analysis\u003c/h3\u003e\n\u003cp\u003eManual segmentation of HCC lesions was conducted to construct the radiomics dataset. Two radiologists, each with more than five years of experience in abdominal MRI interpretation, independently delineated the region of interest (ROI) of HCC tumors on MRI. Subsequently, a third senior radiologist with over ten years of experience reviewed and verified these ROIs to ensure accuracy and consistency. Manual segmentation was performed using ITK-SNAP (version 3.6.0; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.itksnap.org\u003c/span\u003e\u003cspan address=\"http://www.itksnap.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) software.\u003c/p\u003e\u003cp\u003eThe delineations were based on the visible tumor boundaries in MRI sequences, including T2WI, T1WI, DWI, AP, PVP, DP. The final consensus ROIs were subsequently used for radiomics feature extraction and statistical analysis.\u003c/p\u003e\u003cp\u003eImportantly, no qualitative assessments assessed by radiologists were included in this study. All data inputs for model development were based solely on objective information, including clinical laboratory parameters and automatically extracted radiomics features from manually delineated ROIs. This approach minimizes potential interobserver variability and enhances the objectivity, the reproducibility, and the generalizability of the predictive model.\u003c/p\u003e\n\u003ch3\u003e4. Pathological evaluation\u003c/h3\u003e\n\u003cp\u003eAll hematoxylin and eosin (H\u0026amp;E)-stained sections were retrospectively reviewed for MVI by two board-certified pathologists, each with more than 10 years of experience in hepatopathology. The pathologists were blinded to all clinical and imaging data during the assessment. In cases of disagreement, a senior pathologist with over 20 years of diagnostic experience was consulted to reach a consensus.\u003c/p\u003e\u003cp\u003eMVI status was determined according to standardized pathological criteria. Patients were classified into three groups: M0 (no MVI), M1 (1\u0026ndash;5 MVI sites within \u0026le;\u0026thinsp;1 cm from the tumor margin), and M2 (\u0026gt;\u0026thinsp;5 MVI sites or any MVI detected beyond 1 cm from the tumor margin). This grading system has been widely adopted in clinical research for prognostic stratification of HCC[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003e5. Radiomic Feature Extraction\u003c/h3\u003e\n\u003cp\u003eRadiomic feature extraction was conducted following manual segmentation of the region of interest (ROI) by an experienced radiologist using ITK-SNAP software (version 3.6.0)[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The delineated ROIs were used to extract quantitative features with the PyRadiomics library. A total of 1,500 radiomic features were obtained, including first-order statistical descriptors (such as mean, skewness, and kurtosis), shape-based metrics (such as sphericity and elongation), and texture features derived from gray-level matrices, including the Gray-Level Co-occurrence Matrix (GLCM) and the Gray-Level Run-Length Matrix (GLRLM)[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. To better characterize tumor heterogeneity, features were extracted not only from the original images but also from filtered versions processed via wavelet and Laplacian of Gaussian (LoG) transformations[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003e6. Data Preprocessing\u003c/h3\u003e\n\u003cp\u003ePreprocessing involved two major steps:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMissing Value Handling\u003c/b\u003e: Samples with more than 10% missing values were excluded. For the remaining samples, missing values were imputed separately for each feature using k-nearest neighbor (KNN) imputation (k\u0026thinsp;=\u0026thinsp;3), based on the Euclidean distance in the feature space[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eNormalization\u003c/b\u003e: Radiomic features were normalized using Z-score transformation. The mean and standard deviation were calculated exclusively from the training set and then applied to the validation and test sets to ensure no information leakage during model evaluation[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\n\u003ch3\u003e7. Feature Selection\u003c/h3\u003e\n\u003cp\u003eTo mitigate the risk of overfitting and enhance model generalizability, a two-stage feature selection strategy was employed.\u003c/p\u003e\u003cp\u003eIn the initial screening phase, features were first evaluated for reproducibility using the intraclass correlation coefficient (ICC)[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], and only those with ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.75 were retained. Subsequently, univariate statistical tests (t-test or Mann-Whitney U test, depending on data distribution) were performed to exclude features with weak discriminative ability.\u003c/p\u003e\u003cp\u003eIn the dimensionality reduction phase, Recursive Feature Elimination (RFE)[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] based on Random Forest classifiers was conducted to further select the most informative subset of features.\u003c/p\u003e\u003cp\u003eTo systematically evaluate the impact of different feature selection strategies, sixteen pipelines (4 initial screening methods \u0026times; 4 dimensionality reduction methods) were constructed and compared using cross-validation performance within the training set.\u003c/p\u003e\u003cp\u003eUltimately, the pipeline combining \u0026ldquo;t-test or Mann-Whitney U test\u0026thinsp;+\u0026thinsp;RFE with Random Forest\u0026rdquo; was selected for final model training, as it achieved the best balance between feature compactness and model performance.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003e8. Model Architecture and Training\u003c/h3\u003e\n\u003cp\u003eWe developed a Transformer-based deep learning model tailored for three-class MVI prediction. The model architecture was adapted from the Transformer framework, originally introduced for natural language processing tasks[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], and later extended to structured tabular data such as clinical and radiomic features[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the model architecture comprised three main components: an input embedding layer that projected the input features into an 8-dimensional latent space; two stacked Transformer encoder layers, each utilizing 2 attention heads, a feedforward dimension of 16, and a dropout rate of 0.2; and a final output classification layer that generated predictions across three categories.\u003c/p\u003e\u003cp\u003eFollowing the definition of the model architecture, we proceeded to model training. Cross-entropy loss with class-balanced weights was employed to mitigate class imbalance. Model parameters were optimized using the AdamW optimizer with a learning rate of 0.0005 and a weight decay of 0.02. The mini-batch size was set to 16. Training was conducted for 2000 epochs on an NVIDIA RTX 4090 GPU.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003e9. Data Augmentation and Balancing\u003c/h3\u003e\n\u003cp\u003eBased on the extracted radiomic features, the overall preprocessing and modeling strategy is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. This includes feature normalization, missing value imputation, two-step feature selection, and Transformer-based classification. To mitigate the issue of class imbalance in the training set, a hybrid resampling strategy was employed. Specifically, the Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN) was utilized to simultaneously oversample minority classes and clean borderline or noisy samples[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Following resampling, updated class distributions were computed, and class weights were recalculated and incorporated into the cross-entropy loss function to further mitigate residual imbalance during model training.\u003c/p\u003e\u003cp\u003eThis combined approach has been demonstrated to enhance classifier performance in highly imbalanced multiclass medical datasets.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eModel Performance Across Datasets\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe proposed Transformer-based model demonstrated robust classification performance across training, test, and validation sets.\u003c/p\u003e\u003cp\u003eOn the training set, the model achieved a loss of 0.540, an accuracy of 0.759, a weighted F1-score of 0.766, and a macro-average AUC of 0.920.\u003c/p\u003e\u003cp\u003eOn the independent test set, the model yielded a loss of 0.816, an accuracy of 0.733, a weighted F1-score of 0.733, and a macro-average AUC of 0.880 (95% CI: 0.807\u0026ndash;0.953).\u003c/p\u003e\u003cp\u003eOn the external validation set, it reached a loss of 0.948, an accuracy of 0.758, a weighted F1-score of 0.768, and a macro-average AUC of 0.886 (95% CI: 0.833\u0026ndash;0.940).\u003c/p\u003e\u003cp\u003eThe confusion matrices and corresponding ROC curves for all three datasets are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eThe model consistently demonstrated strong discriminative ability for Class 0 across all sets, while Class 1 and Class 2 showed moderate variation in precision and recall.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eSummary of Performance Metrics\u003c/b\u003e\u003c/p\u003e\u003cp\u003eA comprehensive summary of model performance across the training, validation, and test sets is provided in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eOverall performance of the proposed Transformer-based model.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDataset\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eWeighted F1-score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMacro-average Precision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eMacro-average Recall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eMacro-average AUC (95% CI)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTraining Set\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.759\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.766\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.761\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.794\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.920\u003c/p\u003e\u003cp\u003e(0.884\u0026ndash;0.955)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eValidation Set\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.758\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.769\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.676\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.719\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.886\u003c/p\u003e\u003cp\u003e(0.833\u0026ndash;0.940)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTest Set\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.733\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.733\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.635\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.635\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.880\u003c/p\u003e\u003cp\u003e(0.807\u0026ndash;0.953)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eAblation Study with Traditional Machine Learning Models\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo assess the effectiveness of the Transformer-based model, we compared its performance against several traditional machine learning classifiers, including Random Forest, Logistic Regression, XGBoost, and LightGBM. As summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the Transformer-based model achieved the highest accuracy, F1-score, and AUC on both the validation and test sets, outperforming all traditional models across all evaluation metrics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u0026nbsp;\u003c/strong\u003eComparison of the proposed Transformer-based model with traditional machine learning models.\u003c/p\u003e\n\u003cdiv style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\n \u003ctable style=\"border: none;width:446.55pt;\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"border-top:solid windowtext 1.0pt;border-left:none;border-bottom: solid windowtext 1.0pt;border-right:none;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eModel\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:57.45pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eValidation Accuracy\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:62.25pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eValidation F1-score\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:69.35pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eValidation AUC\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:49.7pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eTest Accuracy\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.2pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eTest\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eF1-score\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:68.65pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eTest\u0026nbsp;\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eAUC\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eRandom Forest\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:57.45pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.682\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:62.25pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.694\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:69.35pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.834\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:49.7pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.640\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.2pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.660\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:68.65pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.827\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eLogistic Regression\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:57.45pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.629\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:62.25pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.638\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:69.35pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.837\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:49.7pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.547\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.2pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.575\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:68.65pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.802\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eXGBoost\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:57.45pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.545\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:62.25pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.568\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:69.35pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.807\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:49.7pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.533\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.2pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.557\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:68.65pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.804\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003eLightGBM\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:57.45pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.523\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:62.25pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.547\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:69.35pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.800\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:49.7pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.573\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.2pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.596\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:68.65pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;'\u003e0.823\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"border:none;border-bottom:solid windowtext 1.0pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;color:red;'\u003eTransformer (ours)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:57.45pt;border:none;border-bottom:solid windowtext 1.0pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;color:red;'\u003e0.758\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:62.25pt;border:none;border-bottom:solid windowtext 1.0pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;color:red;'\u003e0.769\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:69.35pt;border:none;border-bottom:solid windowtext 1.0pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;color:red;'\u003e0.886\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:49.7pt;border:none;border-bottom:solid windowtext 1.0pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;color:red;'\u003e0.733\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.2pt;border:none;border-bottom:solid windowtext 1.0pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;color:red;'\u003e0.733\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:68.65pt;border:none;border-bottom:solid windowtext 1.0pt;padding:.75pt .75pt .75pt .75pt;\"\u003e\n \u003cp style='margin:0in;text-align:justify;font-size:14px;font-family:\"Calibri\",sans-serif;'\u003e\u003cstrong\u003e\u003cspan style='font-family:\"Times New Roman\",serif;color:red;'\u003e0.880\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur adoption of a Transformer-based model in this context constitutes a methodological innovation in MVI classification. While Transformer architecture has been increasingly applied in medical imaging, their application to structured radiomic data for multi-class classification remains underrepresented. By tailoring the model to capture dependencies among quantitative features, our framework extends beyond prior binary MVI classification attempts, offering a more nuanced and clinically useful risk stratification strategy.\u003c/p\u003e\u003cp\u003eThis study demonstrated the feasibility and effectiveness of using a Transformer-based deep learning model to perform direct three-class classification of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) patients by integrating radiomic features and clinical laboratory indicators. Compared with conventional machine learning methods such as Random Forest and Logistic Regression, the proposed model consistently achieved superior performance, with a test set AUC of 0.880 (95% CI: 0.807\u0026ndash;0.953) and a weighted F1-score of 0.733. These findings highlight the Transformer architecture\u0026rsquo;s capacity to model complex, high-dimensional, and structured medical data by capturing nonlinear and hierarchical feature dependencies through its self-attention mechanism[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e\u003cp\u003ePrevious radiomics-based studies have primarily focused on binary MVI classification (presence vs. absence), typically reporting AUCs ranging from 0.75 to 0.85[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. In contrast, our approach enables one-step, three-class prediction (M0, M1, M2), providing a more granular and clinically actionable preoperative risk stratification strategy. This advancement allows for refined surgical planning and supports individualized treatment decisions. Additionally, unlike earlier studies that often relied on subjective imaging features manually assessed by radiologists, our model is entirely based on objectively extracted features, which reduces interobserver variability and improves reproducibility[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe key contributions of this study lie in three aspects: proposing a Transformer-based framework for direct MVI three-class classification, demonstrating the benefit of multimodal feature integration from radiomics and clinical data, and ensuring high model reproducibility by using only objective quantitative inputs. Nevertheless, several limitations must be acknowledged. First, although the model performed well on both internal and external validation sets, the dataset was collected from a limited geographic region, which may restrict its generalizability to broader populations and imaging protocols[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Second, the proportion of M2 cases remained low (~\u0026thinsp;15%), and although SMOTE-ENN[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] and class-weighted loss were applied, the class imbalance still posed challenges in reliably distinguishing M1 from M2. Third, as with most deep learning models, the Transformer lacks interpretability, which may hinder its acceptance in clinical decision-making[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eFuture research should focus on external validation using multi-center and multi-ethnic datasets to improve robustness and generalizability[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. In addition, incorporating multi-omics data (e.g., genomics, transcriptomics, proteomics) may further enhance the model\u0026rsquo;s predictive capacity and biological interpretability[\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. To address the black-box nature of the Transformer, explainable artificial intelligence (XAI) techniques\u0026mdash;such as SHAP value analysis, attention heatmaps, and counterfactual reasoning\u0026mdash;should be integrated[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Exploring advanced Transformer variants, including sparse attention mechanisms or hybrid CNN-Transformer architectures, may also contribute to improved efficiency and clinical utility[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this study, we proposed a Transformer-based deep learning framework that integrates radiomic features and clinical laboratory indicators for the preoperative three-class classification of MVI in patients with HCC.\u003c/p\u003e\u003cp\u003eThe proposed model effectively captured complex high-dimensional feature relationships and achieved competitive performance across training, validation, and independent test sets.\u003c/p\u003e\u003cp\u003eBy enabling direct three-class MVI prediction in a single step, our approach provides more granular prognostic information and holds promise for supporting individualized treatment planning and streamlining clinical decision-making.\u003c/p\u003e\u003cp\u003eDespite the encouraging results, further work is needed to enhance the model\u0026rsquo;s clinical applicability.\u003c/p\u003e\u003cp\u003eFuture research should focus on expanding validation cohorts to larger, multi-center, and multi-ethnic datasets, improving model interpretability through explainable artificial intelligence (XAI) techniques, and incorporating multi-omics data to enrich biological insights and further boost predictive accuracy.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eHCC\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eHepatocellular carcinoma\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eMVI\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eMicrovascular invasion\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eMRI\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eMagnetic resonance imaging\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eArea under the curve\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eGd-BOPTA\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eGadolinium benzyloxypropionictetraacetate\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eRFE\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eRecursive feature elimination\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJN and WL conceived and designed the study. CH was responsible for data collection, and contributed to ROI segmentation and radiomics processing along with ZW, GM, and ZX. HR developed the model, performed the data analysis, and drafted the manuscript. JN and WL provided supervision and revised the manuscript critically for important intellectual content. All authors read and approved the final version of the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data used in this study were obtained from the Eastern Hepatobiliary Surgery Hospital and are not publicly available due to institutional and ethical restrictions. Access to the data can be requested from the corresponding author and may be granted upon reasonable request and approval from the institutional review board.\u003c/p\u003e\u003ch2\u003eCritical Relevance Statement\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eAccurate preoperative three-class prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) remains unmet in clinical imaging. This study demonstrates the potential of a Transformer-based model for non-invasive and individualized MVI stratification. \u0026nbsp;\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eH. Sung \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries,\u0026rdquo; \u003cem\u003eCA. Cancer J. Clin.\u003c/em\u003e, vol. 71, no. 3, pp. 209\u0026ndash;249, May 2021, doi: 10.3322/caac.21660.\u003c/li\u003e\n\u003cli\u003eJ. Calderaro, T. P. Seraphin, T. Luedde, and T. G. Simon, \u0026ldquo;Artificial intelligence for the prevention and clinical management of hepatocellular carcinoma,\u0026rdquo; \u003cem\u003eJ. Hepatol.\u003c/em\u003e, vol. 76, no. 6, pp. 1348\u0026ndash;1361, Jun. 2022, doi: 10.1016/j.jhep.2022.01.014.\u003c/li\u003e\n\u003cli\u003eZ. Lei \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Nomogram for Preoperative Estimation of Microvascular Invasion Risk in Hepatitis B Virus\u0026ndash;Related Hepatocellular Carcinoma Within the Milan Criteria,\u0026rdquo; \u003cem\u003eJAMA Surg.\u003c/em\u003e, vol. 151, no. 4, p. 356, Apr. 2016, doi: 10.1001/jamasurg.2015.4257.\u003c/li\u003e\n\u003cli\u003eB. Kim, J. Kahn, and N. A. Terrault, \u0026ldquo;Liver transplantation as therapy for hepatocellular carcinoma,\u0026rdquo; \u003cem\u003eLiver Int.\u003c/em\u003e, vol. 40, no. S1, pp. 116\u0026ndash;121, Feb. 2020, doi: 10.1111/liv.14346.\u003c/li\u003e\n\u003cli\u003eM. Tampaki, G. V. Papatheodoridis, and E. Cholongitas, \u0026ldquo;Intrahepatic recurrence of hepatocellular carcinoma after resection: an update,\u0026rdquo; \u003cem\u003eClin. J. Gastroenterol.\u003c/em\u003e, vol. 14, no. 3, pp. 699\u0026ndash;713, Jun. 2021, doi: 10.1007/s12328-021-01394-7.\u003c/li\u003e\n\u003cli\u003eS. Sumie \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;The Significance of Classifying Microvascular Invasion in Patients with Hepatocellular Carcinoma,\u0026rdquo; \u003cem\u003eAnn. Surg. Oncol.\u003c/em\u003e, vol. 21, no. 3, pp. 1002\u0026ndash;1009, Mar. 2014, doi: 10.1245/s10434-013-3376-9.\u003c/li\u003e\n\u003cli\u003eY. Cai \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Stathmin 1 is a biomarker for diagnosis of microvascular invasion to predict prognosis of early hepatocellular carcinoma,\u0026rdquo; \u003cem\u003eCell Death Dis.\u003c/em\u003e, vol. 13, no. 2, p. 176, Feb. 2022, doi: 10.1038/s41419-022-04625-y.\u003c/li\u003e\n\u003cli\u003eM. Rodr\u0026iacute;guez-Per\u0026aacute;lvarez, T. V. Luong, L. Andreana, T. Meyer, A. P. Dhillon, and A. K. Burroughs, \u0026ldquo;A Systematic Review of Microvascular Invasion in Hepatocellular Carcinoma: Diagnostic and Prognostic Variability,\u0026rdquo; \u003cem\u003eAnn. Surg. Oncol.\u003c/em\u003e, vol. 20, no. 1, pp. 325\u0026ndash;339, Jan. 2013, doi: 10.1245/s10434-012-2513-1.\u003c/li\u003e\n\u003cli\u003eW. Jian \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Predicting microvascular invasion in solitary hepatocellular carcinoma: a multi-center study integrating clinical, MRI assessments, and radiomics indicators,\u0026rdquo; \u003cem\u003eFront. Oncol.\u003c/em\u003e, vol. 15, p. 1511260, Feb. 2025, doi: 10.3389/fonc.2025.1511260.\u003c/li\u003e\n\u003cli\u003eG. Hu \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Radiogenomics nomogram based on MRI and microRNAs to predict microvascular invasion of hepatocellular carcinoma,\u0026rdquo; \u003cem\u003eFront. Oncol.\u003c/em\u003e, vol. 14, p. 1371432, Jul. 2024, doi: 10.3389/fonc.2024.1371432.\u003c/li\u003e\n\u003cli\u003eJ. Li \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;A clinical study exploring the prediction of microvascular invasion in hepatocellular carcinoma through the use of combined enhanced CT and MRI radiomics,\u0026rdquo; \u003cem\u003ePLOS ONE\u003c/em\u003e, vol. 20, no. 1, p. e0318232, Jan. 2025, doi: 10.1371/journal.pone.0318232.\u003c/li\u003e\n\u003cli\u003eP. Lambin \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Radiomics: Extracting more information from medical images using advanced feature analysis,\u0026rdquo; \u003cem\u003eEur. J. Cancer\u003c/em\u003e, vol. 48, no. 4, pp. 441\u0026ndash;446, Mar. 2012, doi: 10.1016/j.ejca.2011.11.036.\u003c/li\u003e\n\u003cli\u003eC. Parmar, P. Grossmann, J. Bussink, P. Lambin, and H. J. W. L. Aerts, \u0026ldquo;Machine Learning methods for Quantitative Radiomic Biomarkers,\u0026rdquo; \u003cem\u003eSci. Rep.\u003c/em\u003e, vol. 5, no. 1, p. 13087, Aug. 2015, doi: 10.1038/srep13087.\u003c/li\u003e\n\u003cli\u003eA. Vaswani \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Attention Is All You Need,\u0026rdquo; Aug. 01, 2023, \u003cem\u003earXiv\u003c/em\u003e: arXiv:1706.03762. Accessed: Nov. 06, 2023. [Online]. Available: http://arxiv.org/abs/1706.03762\u003c/li\u003e\n\u003cli\u003eX. Huang, A. Khetan, M. Cvitkovic, and Z. Karnin, \u0026ldquo;TabTransformer: Tabular Data Modeling Using Contextual Embeddings,\u0026rdquo; Dec. 11, 2020, \u003cem\u003earXiv\u003c/em\u003e: arXiv:2012.06678. doi: 10.48550/arXiv.2012.06678.\u003c/li\u003e\n\u003cli\u003efor the Liver Cancer Pathology Group of China (LCPGC) \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;A standardized pathological proposal for evaluating microvascular invasion of hepatocellular carcinoma: a multicenter study by LCPGC,\u0026rdquo; \u003cem\u003eHepatol. Int.\u003c/em\u003e, vol. 14, no. 6, pp. 1034\u0026ndash;1047, Dec. 2020, doi: 10.1007/s12072-020-10111-4.\u003c/li\u003e\n\u003cli\u003eP. A. Yushkevich \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability,\u0026rdquo; \u003cem\u003eNeuroImage\u003c/em\u003e, vol. 31, no. 3, pp. 1116\u0026ndash;1128, Jul. 2006, doi: 10.1016/j.neuroimage.2006.01.015.\u003c/li\u003e\n\u003cli\u003eJ. J. M. Van Griethuysen \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Computational Radiomics System to Decode the Radiographic Phenotype,\u0026rdquo; \u003cem\u003eCancer Res.\u003c/em\u003e, vol. 77, no. 21, pp. e104\u0026ndash;e107, Nov. 2017, doi: 10.1158/0008-5472.CAN-17-0339.\u003c/li\u003e\n\u003cli\u003eA. Zwanenburg \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping,\u0026rdquo; \u003cem\u003eRadiology\u003c/em\u003e, vol. 295, no. 2, pp. 328\u0026ndash;338, May 2020, doi: 10.1148/radiol.2020191145.\u003c/li\u003e\n\u003cli\u003eC. Parmar \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Robust Radiomics Feature Quantification Using Semiautomatic Volumetric Segmentation,\u0026rdquo; \u003cem\u003ePLoS ONE\u003c/em\u003e, vol. 9, no. 7, p. e102107, Jul. 2014, doi: 10.1371/journal.pone.0102107.\u003c/li\u003e\n\u003cli\u003eO. Troyanskaya \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Missing value estimation methods for DNA microarrays,\u0026rdquo; \u003cem\u003eBioinformatics\u003c/em\u003e, vol. 17, no. 6, pp. 520\u0026ndash;525, Jun. 2001, doi: 10.1093/bioinformatics/17.6.520.\u003c/li\u003e\n\u003cli\u003eG. E. A. P. A. Batista, R. C. Prati, and M. C. Monard, \u0026ldquo;A study of the behavior of several methods for balancing machine learning training data,\u0026rdquo; \u003cem\u003eACM SIGKDD Explor. Newsl.\u003c/em\u003e, vol. 6, no. 1, pp. 20\u0026ndash;29, Jun. 2004, doi: 10.1145/1007730.1007735.\u003c/li\u003e\n\u003cli\u003eC. Sammut and G. I. Webb, Eds., \u003cem\u003eEncyclopedia of Machine Learning and Data Mining\u003c/em\u003e. Boston, MA: Springer US, 2017. doi: 10.1007/978-1-4899-7687-1.\u003c/li\u003e\n\u003cli\u003eI. Guyon, J. Weston, and S. Barnhill, \u0026ldquo;Gene Selection for Cancer Classification using Support Vector Machines\u0026rdquo;.\u003c/li\u003e\n\u003cli\u003eG. Somepalli, M. Goldblum, A. Schwarzschild, C. B. Bruss, and T. Goldstein, \u0026ldquo;SAINT: Improved Neural Networks for Tabular Data via Row Attention and Contrastive Pre-Training,\u0026rdquo; Jun. 02, 2021, \u003cem\u003earXiv\u003c/em\u003e: arXiv:2106.01342. doi: 10.48550/arXiv.2106.01342.\u003c/li\u003e\n\u003cli\u003eS.-T. Feng \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Preoperative prediction of microvascular invasion in hepatocellular cancer: a radiomics model using Gd-EOB-DTPA-enhanced MRI,\u0026rdquo; \u003cem\u003eEur. Radiol.\u003c/em\u003e, vol. 29, no. 9, pp. 4648\u0026ndash;4659, Sep. 2019, doi: 10.1007/s00330-018-5935-8.\u003c/li\u003e\n\u003cli\u003eM. He, P. Zhang, X. Ma, B. He, C. Fang, and F. Jia, \u0026ldquo;Radiomic Feature-Based Predictive Model for Microvascular Invasion in Patients With Hepatocellular Carcinoma,\u0026rdquo; \u003cem\u003eFront. Oncol.\u003c/em\u003e, vol. 10, p. 574228, Nov. 2020, doi: 10.3389/fonc.2020.574228.\u003c/li\u003e\n\u003cli\u003eW. Zhang, Y. Guo, and Q. Jin, \u0026ldquo;Radiomics and Its Feature Selection: A Review,\u0026rdquo; \u003cem\u003eSymmetry\u003c/em\u003e, vol. 15, no. 10, p. 1834, Sep. 2023, doi: 10.3390/sym15101834.\u003c/li\u003e\n\u003cli\u003eR. D. Riley \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;External validation of clinical prediction models using big datasets from e-health records or IPD meta-analysis: opportunities and challenges,\u0026rdquo; \u003cem\u003eBMJ\u003c/em\u003e, p. i3140, Jun. 2016, doi: 10.1136/bmj.i3140.\u003c/li\u003e\n\u003cli\u003eA. Holzinger, C. Biemann, C. S. Pattichis, and D. B. Kell, \u0026ldquo;What do we need to build explainable AI systems for the medical domain?,\u0026rdquo; Dec. 28, 2017, \u003cem\u003earXiv\u003c/em\u003e: arXiv:1712.09923. doi: 10.48550/arXiv.1712.09923.\u003c/li\u003e\n\u003cli\u003eR. D. Riley \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Evaluation of clinical prediction models (part 2): how to undertake an external validation study,\u0026rdquo; \u003cem\u003eBMJ\u003c/em\u003e, p. e074820, Jan. 2024, doi: 10.1136/bmj-2023-074820.\u003c/li\u003e\n\u003cli\u003eP. Kickingereder \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Radiogenomics of Glioblastoma: Machine Learning\u0026ndash;based Classification of Molecular Characteristics by Using Multiparametric and Multiregional MR Imaging Features,\u0026rdquo; \u003cem\u003eRadiology\u003c/em\u003e, vol. 281, no. 3, pp. 907\u0026ndash;918, Dec. 2016, doi: 10.1148/radiol.2016161382.\u003c/li\u003e\n\u003cli\u003eS. M. Lundberg and S.-I. Lee, \u0026ldquo;A Unified Approach to Interpreting Model Predictions\u0026rdquo;.\u003c/li\u003e\n\u003cli\u003eE. Tjoa and C. Guan, \u0026ldquo;A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI,\u0026rdquo; \u003cem\u003eIEEE Trans. Neural Netw. Learn. Syst.\u003c/em\u003e, vol. 32, no. 11, pp. 4793\u0026ndash;4813, Nov. 2021, doi: 10.1109/TNNLS.2020.3027314.\u003c/li\u003e\n\u003cli\u003eK. Choromanski \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Rethinking Attention with Performers,\u0026rdquo; Nov. 19, 2022, \u003cem\u003earXiv\u003c/em\u003e: arXiv:2009.14794. doi: 10.48550/arXiv.2009.14794.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Hepatocellular carcinoma, Microvascular invasion, Transformer model, Radiomics, Deep learning, Multimodal integration","lastPublishedDoi":"10.21203/rs.3.rs-7131534/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7131534/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eMicrovascular invasion (MVI) is a critical prognostic factor in hepatocellular carcinoma (HCC), but preoperative three-class prediction remains challenging. Radiomics and clinical biomarkers may enable more accurate and individualized assessment.\u003c/p\u003e\u003ch2\u003eAim\u003c/h2\u003e\u003cp\u003eTo develop and validate a Transformer-based deep learning framework that integrates radiomic and clinical features for direct three-class MVI classification in HCC patients.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eThis retrospective study included 438 patients with pathologically confirmed HCC and MVI status from a single institution. Radiomic features were extracted from preoperative Gd-BOPTA-enhanced MRI, and clinical laboratory data were collected. A two-stage feature selection strategy, combining univariate statistical testing and recursive feature elimination, was applied. A Transformer-based model was built to classify three MVI categories (M0, M1, M2), and its performance was evaluated on internal and external test sets. Results were compared with traditional machine learning models, including Random Forest, Logistic Regression, XGBoost, and LightGBM.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe Transformer-based model achieved an accuracy of 0.733, a weighted F1-score of 0.733, and a macro-average AUC of 0.880 (95% CI: 0.807\u0026ndash;0.953) on the internal test set. On the external validation set, it reached an accuracy of 0.758, a weighted F1-score of 0.768, and a macro-average AUC of 0.886 (95% CI: 0.833\u0026ndash;0.940). It outperformed traditional classifiers and showed superior ability to identify high-risk M2 cases.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eThis Transformer-based model enables accurate and objective three-class MVI prediction using multimodal features, supporting individualized surgical planning and improved clinical outcomes.\u003c/p\u003e","manuscriptTitle":"Transformer-Based Deep Learning for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-28 11:07:14","doi":"10.21203/rs.3.rs-7131534/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"dee80a2a-0d04-4e20-b2ee-cb013a13a229","owner":[],"postedDate":"July 28th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-08-05T10:53:44+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-28 11:07:14","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7131534","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7131534","identity":"rs-7131534","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}