Artificial Intelligence Model for Accurate Grading and Staging of Chronic Hepatitis B: Development and Validation

Artificial Intelligence Model for Accurate Grading and Staging of Chronic Hepatitis B: Development and Validation | 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 Article Artificial Intelligence Model for Accurate Grading and Staging of Chronic Hepatitis B: Development and Validation Xinyan Zhao, Jialing Zhou, Ying Zhang, Yongyin Li, Lin Wang, Guoxin Teng, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7440962/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 and aims: Chronic hepatitis B (CHB) is a global challenge, with histological assessment affected by observer variability. We aim to develop and validate BJ-HepaGS, an AI model for consistent evaluation. Methods: BJ-HepaGS was developed using hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) from CHB patients across multiple hospitals. Model performance was validated using area under the curve (AUC) and intraclass correlation coefficient (ICC) metrics. Results: The BJ-HepaGS trained on 673 WSIs (627,703 patches) and validated on independent (300 H&E-WSIs, 331,037 patches) and paired cohorts (n = 100 H&E-WSIs, 52,271 patches). The independent set achieved area under the curve (AUC) values of 0.91–0.98 for grading and 0.85–0.91 for staging, with strong consistency versus expert consensus (ICC = 0.824 and 0.681, respectively). BJ-HepaGS distinguished fibrosis stage F0-1 vs. F2-4 (87.6% accuracy) and cirrhosis (F0-3 vs. F4; 86.0% accuracy), and reliably assessed inflammation improvement ( p = 0.885) and fibrosis regression ( p = 0.388) in pre- and post-treatment paired samples. With the assistance of AI, the consistency between senior and junior expert interpretations on inflammation and fibrosis were significantly enhanced (both p < 0.001). Conclusions: BJ-HepaGS addresses a key gap in CHB care by providing reproducible, objective histopathological interpretation, supporting standardized diagnosis and improved clinical management. Health sciences/Medical research/Experimental models of disease Health sciences/Gastroenterology/Gastrointestinal models Chronic viral hepatitis Necroinflammation Fibrosis Cirrhosis METAVIR system Hematoxylin and Eosin (H&E) staining Whole slide images Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Chronic hepatitis B (CHB) remains a major global health challenge, with approximately 254 million people living with chronic HBV infection as of 2022. [1] CHB is responsible for 820,000 to 1.4 million annual deaths worldwide, primarily due to complications such as cirrhosis and hepatocellular carcinoma (HCC). 1 , 2 The burden is particularly pronounced in the Asian Pacific, where prevalence rates are highest, and in low- and middle-income countries, where the economic effect is substantial. 3 , 4 To address this issue, the World Health Organization (WHO) has set ambitious targets to eliminate viral hepatitis as a public health threat by 2030; WHO aims to reduce new HBV infections by 90% and HBV-related deaths by 65% compared to 2015 levels. 5 , 6 Accurate assessment of inflammatory activity (grading) and liver fibrosis (staging) is critical for the effective management of CHB. 7 – 11 Liver biopsy remains the gold standard for grading and staging because it provides detailed histopathological information essential for clinical decision-making. 12 – 14 Reliable histological evaluation also underpins the development of noninvasive models for predicting inflammatory grading and fibrosis staging. However, the clinical utility of liver biopsy is limited by significant inter- and intra-observer variability in histopathological interpretation, which undermines consistent disease assessment and management. 15 – 18 These limitations highlight the urgent need for more objective and reproducible diagnostic approaches to improve patient care. While artificial intelligence (AI) has shown promise in enhancing diagnostic accuracy and standardization in hepatology, its application in liver histopathology has primarily focused on HCC diagnosis. 19 – 22 The majority of AI-driven studies in this field have concentrated on improving the detection and characterization of HCC; to our knowledge, few published AI studies have focused on non-tumor liver diseases such as CHB. This paucity of research underscores a critical unmet need in the field. In this study, we aimed to develop and validate a novel AI-based model to enhance the accuracy and reliability of CHB grading and staging, ultimately contributing to better clinical care for patients with CHB. Patients and Methods Patient Population and Study Design This multi-center, retrospective study enrolled patients with CHB from four Chinese hospitals from January 2013 to December 2024. The study was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (2024-P2-042-01). Written informed consent was waived. All procedures were conducted in accordance with the Declaration of Helsinki and relevant local regulations. A total of 3,219 liver biopsy slides, each including hematoxylin and eosin (H&E), Masson’s trichrome, and reticulin staining, were included (1,023 patients). The slides represented patients with diverse inflammation grades and fibrosis stages, before or after treatment. The inclusion criteria were as follows: (a) age ≥ 18 years, regardless of gender; (b) CHB diagnosis (hepatitis B surface antigen positive for > 6 months); (c) pathological liver biopsy indicated a length > 1.0 cm and/or more than six portal tracts. The exclusion criteria were as follows: (a) co-occurrence of any other liver diseases, including autoimmune hepatitis, alcoholic liver disease, or HCC at the time of liver biopsy; (b) co-infection with any other viral hepatitis, such as hepatitis C or D; and (c) poor staining quality of the liver biopsy samples. Sample Selection and Interpretation Liver biopsy was performed in the right lobe of the liver using a 16- or 18-gauge needle. The specimens were stained at the different centers using H&E, Masson’s trichrome, and reticulum staining to assess necroinflammatory activity and fibrosis. Liver biopsy slides were uniformly scanned using a consistent device (3DHISTECH, Budapest, Hungary) and digitally processed for AI analysis. Interpretation Methods Four experts—three pathologists and one hepatologist with specialized pathological training—were divided into two groups (Senior Group A: L. Wang and X.Y. Zhao; Junior Group B: X.L. Li and G.X. Teng). Each group independently interpreted the whole-slide images (WSIs) for inflammation grading and fibrosis staging according to the METAVIR system. Discrepancies were resolved through re-evaluation to achieve consensus. The consensus interpretation served as the ground truth for training the AI model. Clinical Data Collection Clinical data, collected at the time of pathological examination, included demographic parameters (age and gender) and the following laboratory parameters: platelet count (PLT), alanine aminotransferase (ALT), aspartate aminotransferase (AST), serum albumin (ALB), total bilirubin (TBIL), HBV DNA, and HBeAg status. Development of the AI Model AI-based Analysis Method We employed the clustering-constrained attention multiple instance learning model (CLAM), 23 a robust AI approach for WSI analysis. The CLAM model consists of four key components (Fig. 1 ): Feature Extraction : A pretrained UNI feature extractor encodes image patches into representative features, 24 leveraging its ability to learn diverse medical image patterns. Gated Attention Layer : An attention layer assigns attention scores to image patches, prioritizing regions critical for classification. Clustering Constraint : Pseudo-labels for high- and low-attention patches are generated during training, enhancing supervisory signals and feature representation. Fully Connected Layer : Global feature representations are mapped to class labels, completing the classification task. Dataset Allocation A total of 673 H&E WSIs were used for training. Two external datasets were used to validate the model: one dataset of 300 H&E WSIs from four centers; and a paired liver biopsy dataset of 100 H&E WSIs. These datasets were employed to train and validate the model, ensuring accurate grading and staging of inflammation and fibrosis based on H&E staining. Training Process WSIs were divided into non-overlapping 512×512-pixel patches, separating tissue regions from background pixels. The UNI extractor was applied to extract features, followed by attention score computation via the gated attention layer. Pseudo-labels were generated through clustering constraints, and features were aggregated using attention-weighted representations for the final task. The model incorporates five attention branches to assess tissue patch significance, enhancing interpretability. Performance was evaluated based on the area under the receiver operating characteristic curve (AUROC) and accuracy. Training was conducted using an NVIDIA DGX A100 GPU with an 80-GB memory capacity. Statistical Analysis We used various statistical methods to analyze the data and assess the AI model performance. Continuous variables (e.g., age and biochemical markers) were summarized by median, and interquartile range, while categorical variables (e.g., gender) were described by frequencies and percentages. Categorical comparisons were made using chi-square or Fisher’s exact tests, with two-sided p < 0.05 indicating significance. These analyses were conducted using SPSS (version 26.0) and GraphPad Prism (version 6.02). When comparing AI model performance with expert scores, paired t-tests were applied for continuous variables and chi-square tests for categorical variables. ROC curves were generated using MedCale software (version 15.8) to evaluate the diagnostic accuracy of the AI model for inflammation grading and fibrosis staging. Model generalizability was tested on an independent validation set using metrics of accuracy and specificity. A paired dataset was also used to assess the ability of the model to evaluate dynamic changes of grading and staging. The AI model was developed and optimized using Python (version 3.10.4). Results A total of 1,217 patients were screened for this study. After excluding 194 patients due to unqualified histological samples and missing serological results, 1023 patients, including 3,219 WSIs, were included in the final analysis ( Fig. 2 ). The dataset was divided into three sets: 673 WSIs (627,703 patches) for the training set, 300 WSIs (331,037 patches) for the validation set, and 100 WSIs (52,271 patches) for the paired set. Clinical and Histopathological Characteristics of the Training and Validation Sets A total of 673 WSIs were included in the training dataset to develop the BJ-HepaGS model. The ratio of patients before treatment to those after treatment was approximately 2:1. The cohort was predominantly male (76%). Among the patients, 37% and 31% exhibited HBV DNA and were HBeAg positive, respectively. A total of 463 patients (69%) had no or minimal inflammation and were graded as G0 or G1. Moderate inflammation (G2) was observed in 23% of patients, while severe inflammatory activity (G3/4) was presented in the smallest proportion of patients (8%). Regarding fibrosis staging, 279 (41%) patients were staged at F0 or F1. A total of 77 (11%) patients were staged at F2. Advanced fibrosis accounted for a larger proportion, with 218 (32%) patients at F3 and 99 (15%) at F4. The ratio of patients before treatment to patients after treatment in the independent validation set was approximately 1:5. No significant differences were observed between the training set and validation set 1 in terms of inflammation grade, ALT, AST, ALB, TBIL, PLT, or HBeAg positivity ( p > 0.05). However, a significant difference was observed in the distribution of fibrosis stages ( p < 0.05), which might be attributed to the higher proportion of post-treatment cases in this validation set. The paired validation set included 50 cases (100 H&E WSIs), with both pre-treatment and post-treatment samples available. Significant improvements in both inflammation and fibrosis were observed after treatment. Prior to treatment, the distribution of grades was as follows: 14 patients (28%) graded G0 or G1, 18 (36%) G2, and 18 (36%) G3/4. In terms of fibrosis stages, 19 patients (38%) were staged at F0 or F1, 5 (10%) at F2 , 15 (30%) at F3, and 11 (22%) at F4. The baseline characteristics of all patients were summarized in Table 1 . Visualization of High-attention Patch Selection and Heatmaps Across Inflammation Grades and Fibrosis stages in the BJ-HepaGS Model The overall workflow of the BJ-HepaGS model for analyzing H&E-stained liver biopsy slides was illustrated in Fig. 3A . The WSIs were first divided into fixed-size patches, and features were extracted from each individual patch. The model identified high-attention patches across various stages of inflammation (G1–G4) and fibrosis (F1–F4), as illustrated in Fig. 3B . To further assess the distribution of learned features, we conducted principal component analysis (PCA) on the patch-level embeddings derived from a representative WSI. As illustrated in the bottom panel, patches with higher attention scores (indicated by darker colors) tended to cluster together in the reduced feature space ( Fig. 3C ). The attention-based heatmap visualizations generated by the model over the WSIs are shown in Fig. 3D . Color overlays on each WSI indicate the attention strength assigned by the model, with red denoting high attention and blue indicating low attention. The magnified views highlight a strong correspondence between high-attention regions and histologically relevant features, suggesting that the model effectively localized tissue patterns associated with inflammation and fibrosis. Performance of the BJ-HepaGS Model in Grading and Staging: An Independent Validation Set In the validation set of 300 H&E WSIs, we assessed the performance of the BJ-HepaGS model by evaluating its diagnostic performance and comparing its outputs to the consensus of expert pathologists ( Fig. 4 ). Grading : The BJ-HepaGS model performed well in grading. The model accurately identified inflammation of grades ≥ G1 (AUROC = 0.91; 95% CI: 0.87–0.95), ≥ G2 (AUROC = 0.95; 95% CI: 0.92–0.97), and ≥ G3 (AUROC = 0.98; 95% CI: 0.97–0.99). The model also achieved high accuracy for assessing all grades: ≥ G1 (89.0%), ≥ G2 (87.6%), and ≥ G3 (94.6%) (Table 2) . The confusion matrices also indicated that the model correctly classified grades ≥ G3 (Fig. 5). . Staging : The BJ-HepaGS model demonstrated good performance in staging. In particular, the model excelled in identifying advanced stages: ≥ F2 (AUROC = 0.89; 95% CI: 0.84–0.93), ≥ F3 (AUROC = 0.87; 95% CI: 0.83–0.91), and F4 (AUROC = 0.82; 95% CI: 0.76–0.88). The model also showed good accuracy for assessing all stages: ≥ F2 (85.0), ≥ F3 (81.0%), and F4 (84.7%). (Table 2) The confusion matrices also demonstrated that the model correctly classified stages ≥ F2 ( Fig. 5 ). Concordance of the BJ-HepaGS Model with Expert Consensus in Grading and Staging In the validation set of 300 H&E WSIs and the paired validation set of 100 H&E WSIs, we assessed the concordance of the BJ-HepaGS model and expert consensus of ICC ( Table 3 ). Grading : The model demonstrated strong agreement with the expert consensus for inflammation grading, achieving an ICC of 0.824 (95% CI: 0.784–0.857). The performance of the model remained consistent across different grades: G ≥ 1 (ICC = 0.721; 95% CI: 0.662–0.771), G ≥ 2 (ICC = 0.721; 95% CI: 0.661–0.771), and G ≥ 3 (ICC = 0.712; 95% CI: 0.652–0.764). In the analysis of 50 paired samples, the BJ-HepaGS model achieved an ICC of 0.865 (95% CI: 0.805–0.907). The AI model assessments of inflammation were closely aligned with the consensus of human experts ( Table 2 ). Staging : The model also achieved a moderate ICC of 0.681 (95% CI: 0.615–0.737) for staging. It showed consistency for fibrosis staging: F ≥ 2 (ICC = 0.644; 95% CI: 0.573–0.706), F ≥ 3 (ICC = 0.613; 95% CI: 0.537–0.679) and F4 (ICC = 0.369; 95% CI: 0.267–0.463). In the 50 paired samples, the model also demonstrated moderate consistency for fibrosis staging, with an ICC of 0.682 (95% CI: 0.562–0.775). Assessment of inflammation improvement and fibrosis regression in paired pre- and post-treatment samples Inflammation improvement was defined as a decrease of 1 or more in the inflammation grade; stable inflammation was defined as no change; and worsening inflammation was defined as an increase of 1 or more in grade. According to the BJ-HepaGS model, 33 cases (66%) demonstrated improvement, 13 cases (26%) remained stable, and 4 cases (8%) exhibited worsening. Based on expert consensus, 34 cases (68%) showed improvement, 11 cases (22%) remained stable, and 5 cases (10%) worsened. As for the change in inflammation, the distribution of outcomes between the AI model and expert consensus was consistent, with no statistically significant difference ( p = 0.885; Fig. 4 and Supplementary Table 1 ). Fibrosis regression was defined as a decrease of 1 or more in the fibrosis stage; stability was defined as no change; and progression was defined as an increase of 1 or more. The BJ-HepaGS model identified fibrosis regression in 22 cases (44%), stability in 25 cases (50%), and progression in 3 cases (6%). Expert consensus found regression in 23 cases (46%), stability in 23 cases (46%), and progression in 4 cases (8%). As for the change in fibrosis, the AI model and expert consensus was consistent with no statistically significant difference ( p = 0.388; Fig. 4 and Supplementary Table 1 ). These results demonstrate that the BJ-HepaGS model is capable of interpreting dynamic changes in CHB, supporting its potential as a reliable tool for evaluating treatment response and monitoring disease progression. Enhanced Concordance with Expert Interpretations After Iterative Refinement In the validation set of 300 cases, discrepancies between the BJ-HepaGS model and expert consensus were observed in approximately 20% of inflammation assessments and 40% of fibrosis assessments. Notably, the AI model provided substantial assistance to both the senior and junior expert groups during re-evaluation ( Fig. 4 and Supplementary Table 2 ). Grading : After reviewing the suggestions from the AI model, 44% (26/59) of senior experts and 46% (32/70) of junior experts revised their grading scores. Among those who changed their scores, most of the senior experts (88%, 23/26) and junior experts (88%, 28/32) aligned their revised scores with the AI model; only a small minority adjusted their scores to alternative scores. Staging : After reviewing the suggestions from the AI model, 29% (34/117) of senior experts and 31% (40/128) of junior experts revised their scores. Among those who changed their scores, most of the senior experts (76%, 26/34) and an even higher proportion of the junior experts (80%, 32/40) aligned their revised scores with the AI model; only a small minority adjusted their scores to alternative scores. Iterative AI assistance improved agreement between senior and junior expert scores, resulting in a model closely matching expert grading. For inflammation, the ICC increased from 0.730 (95% CI: 0.673-0.779) to 0.857 (95% CI: 0.824–0.884). For fibrosis, the ICC increased from 0.852 [95% CI: 0.817–0.880) to 0.906 [95% CI: 0.883–0.924]. Concordance remained high in cirrhosis (ICC 0.822, 95% CI: 0.782–0.856; Table 3). These findings demonstrate that the AI model substantially improved the precision and reliability of expert interpretations, enabling more accurate liver biopsy assessments (Table 4). Discussion The BJ-HepaGS model, based on only H&E-stained WSIs, demonstrated expert-level accuracy in grading and staging CHB. The model achieved an AUROC of 0.91 for necroinflammation grading and 0.81 for fibrosis staging, indicating high performance relative to expert consensus. Furthermore, in the paired validation set, the model exhibited significant correlations with expert assessments of changes in necroinflammation grade ( p = 0.855) and fibrosis stage ( p = 0.388). In approximately one-third of the cases where discrepancies existed between the AI model and the scores of different groups of liver pathology experts, the implementation of the BJ-HepaGS model facilitated modifications in the expert scores, underscoring its capacity to reduce both inter-observer and intra-observer variability. Liver biopsy remains the gold standard for assessing CHB. Liver biopsy serves as the foundation for noninvasive model development, treatment decision-making, and outcome prediction. However, histopathological assessment is subject to substantial inter-observer variability. 25 – 27 As the first AI model specifically designed to provide objective and reproducible assessments, the BJ-HepaGS model represents significant progress toward addressing this limitation. Notably, the model achieves near-expert consistency, with an ICC of 0.824 for inflammation grading and 0.68 for fibrosis staging, demonstrating its potential to reduce variability in CHB assessment. Currently, machine learning models are primarily implemented in various pathology applications, including cancer detection, tumor origin identification, and metabolic dysfunction-associated steatohepatitis (MASH). 28 – 30 Sanyal et al. recently demonstrated the strong performance of AIM-MASH (AI-based measurement of MASH) in histological assessments, achieving kappa values of 0.67 (95% CI: 0.64–0.71) for lobular inflammation and 0.62 (95% CI: 0.58–0.65) for fibrosis. 31 Notably, the performance of the AIM-MASH tool is comparable to that of our BJ-HepaGS model, highlighting the potential of deep learning in pathological assessments. By leveraging advanced deep learning techniques, these AI-based models can enhance the precision and consistency of histological evaluations while minimizing the subjectivity and variability inherent in manual scoring systems. The advanced design of our model enables it to focus on diagnostically relevant regions within H&E-stained WSIs, thereby enhancing both interpretability and accuracy. The AIM-MASH system analyzes WSIs using both H&E and Masson’s trichrome staining and employs convolutional neural networks to generate color-coded overlays for histological feature segmentation. 29 In contrast, our model operates on routine H&E-stained slides alone; thus, by eliminating the need for special stains like Sirius Red, our model improves clinical practicality and scalability. These technical advantages position the BJ-HepaGS model as a robust and deployable solution for AI-assisted liver pathology. A major strength of this study is the use of a large, diverse, and multi-center dataset, which enhanced the generalizability of the BJ-HepaGS model. Standardized histopathological criteria were applied throughout, ensuring consistency and reliability in both model training and evaluation. The performance of the model was rigorously benchmarked against expert consensus, demonstrating robust accuracy in cross-sectional grading and staging as well as in the longitudinal assessment of histological changes, an area where many existing models fall short due to limited validation or lack of dynamic assessment. This study has several limitations. While the model performed well in inflammation grading, its accuracy for fibrosis staging was lower, likely due to the challenge of detecting subtle changes on H&E stains alone. Expanding the training set with more severe fibrosis cases and incorporating fibrosis-specific stains, such as Masson’s trichrome, may improve performance. Additionally, since most patients were from Chinese hospitals, results may not be generalizable to other populations. Successful integration of BJ-HepaGS into clinical practice will require ensuring compatibility with digital pathology systems, effective user training, and regular model updates as guidelines evolve. Prospective, real-world studies are needed to validate its impact on clinical decision-making and patient outcomes. Future research should also explore combining AI-based histology with clinical and laboratory data to provide comprehensive decision support for CHB management. Conclusion In summary, the BJ-HepaGS model represents a significant advancement in AI-assisted liver pathology. The model offers expert-level accuracy, reproducibility, and the ability to monitor dynamic histological changes. By addressing key challenges in the assessment of liver biopsy slides, this model has the potential to standardize care, improve diagnostic confidence, and ultimately enhance outcomes for patients with CHB. Abbreviations CHB, chronic hepatitis B; HBV, hepatitis B virus; AI, artificial intelligence; H&E, hematoxylin and eosin; WSIs, whole-slide images; CLAM, clustering-constrained attention multiple instance learning model; ALB, albumin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; TBIL, total bilirubin; AUROC, area under the receiver operating characteristic curve; HBeAg, Hepatitis B e antigen positive; LSM, liver stiffness measurement; PLT, platelet count; ROC, receiver operating characteristic curve; ULN, upper limit of normal; SD, standard deviation; IQR, interquartile range; CI, confidence interval; ICC, Intraclass Correlation Coefficient. Declarations Author s’ c ontribution: J.L. Zhou, Y Zhang and Y.Y. Li contribute equally to the study. X.Y. Zhao designed the study, with assistance from J.L. Zhou. Y.Y. Li, S. Yang, H. Liu collected the data. X.Y. Zhao, L Wang, G.X. Teng and X.L. Li interpret the inflammation and fibrosis. C.H. Hu and Y. Zhang trained the model. J.L. Zhou prepared the graphs and drafted the manuscript. X.Y. Zhao supervised the work, reviewed the manuscript, and provided critical insights during the editing process. All co-authors reviewed, revised, and approved the final version of the manuscript. Data Transparency Statement: Data will be made available on request. C onflict of interest statements : The authors disclose no conflicts. F inancial support statement : This work was supported by the National Key Research and Development Program of China (No.2022YFC2303600), the National key clinical specialist construction Programs (2022001) and the National Natural Science Foundation of China under Grants No. 82371960. Acknowledgments : We thank all individuals and patients who participated in this study. We also thank all pathologists who participated in the study. We thank AiMi Academic Services (www.aimieditor.com) for English language editing and review services. Code availability The BJ-HepaGS code is available at https://github.com/zznn-ying/BJ-HepaGS.git References Burki T. WHO's 2024 global hepatitis report. Lancet Infect Dis. 2024; 24: e362-e363. doi: 10.1016/S1473-3099(24)00307-4. Jeng WJ, Papatheodoridis GV, Lok ASF. Hepatitis B. Lancet. 2023; 401: 1039-1052. doi: 10.1016/S0140-6736(22)01468-4. World Health Organization. Global progress report on HIV, viral hepatitis and sexually transmitted infections, 2021. Accountability for the global health sector strategies 2016-2021: actions for impact[EB/OL]. 2021. https://www.who.int/publications/i/item/9789240027077. Hsu, YC., Huang, D.Q. & Nguyen, M.H. Global burden of hepatitis B virus: current status, missed opportunities and a call for action. Nat Rev Gastroenterol Hepatol. 2023; 20: 524-537. doi: 10.1038/s41575-023-00760-9. McMahon BJ. Meeting the WHO and US Goals to Eliminate Hepatitis B Infection by 2030: Opportunities and Challenges. Clin Liver Dis (Hoboken). 2018; 12: 29-32. doi: 10.1002/cld.733. World Health Organization. Global health sector strategy on viral hepatitis 2016- 2021: Towards ending viral hepatitis[EB/OL]. https://apps.who.int/iris/bitstream/10665/246177/1/WHO-HIV-2016.06-eng,pdf?ua= 1. https://apps.who.int/iris/bitstream/10665/246177/1/WHO-HIV-2016.06-eng Zhang X, Jie Y, Wan Z, et al. Prognostic Value of Inflammatory Indicators in Chronic Hepatitis B Patients With Significant Liver Fibrosis: A Multicenter Study in China. Front Pharmacol. 2021; 12: 653751. doi: 10.3389/fphar.2021.653751. Bera C, Hamdan-Perez N, Patel K. Non-Invasive Assessment of Liver Fibrosis in Hepatitis B Patients. J Clin Med. 2024; 13 :1046. doi: 10.3390/jcm13041046. Terrault NA, Lok ASF, McMahon BJ, et al. Update on prevention, diagnosis, and treatment of chronic hepatitis B: AASLD 2018 hepatitis B guidance. Hepatology. 2018; 67: 1560-1599. doi: 10.1002/hep.29800. European Association for the Study of the Liver. EASL 2017 Clinical Practice Guidelines on the management of hepatitis B virus infection. J Hepatol. 2017; 67: 370-398. doi: 10.1016/j.jhep.2017.03.021. World Health Organization. Guidelines for the prevention,diagnosis, care and treatment for people with chronic hepatitis B infection[EB/OL]. (2024-03-29)[2024-12-18]. https://www.who.int/publications/i/item/9789240090903. Mani H, Kleiner DE. Liver biopsy findings in chronic hepatitis B. Hepatology. 2009; 49: S61-71. doi: 10.1002/hep.22930. Tong XF, Wang QY, Zhao XY, et al. Histological assessment based on liver biopsy: the value and challenges in NASH drug development. Acta Pharmacol Sin. 2022; 43: 1200-1209. doi: 10.1038/s41401-022-00874-x. Castera L, Hartmann DJ, Chapel F, et al. Serum laminin and type IV collagen are accurate markers of histologically severe alcoholic hepatitis in patients with cirrhosis. J Hepatol. 2000; 32: 412-418. doi: 10.1016/s0168-8278(00)80391-8. Kishanifarahani Z, Ahadi M, Kazeminejad B, et al. Inter-observer Variability in Histomorphological Evaluation of Non-neoplastic Liver Biopsy Tissue and Impact of Clinical Information on Final Diagnosis in Shahid Beheshti University of Medical Sciences Affiliated Hospitals. Iran J Pathol. 2019; 14: 243-247. doi: 10.30699/ijp.2019.99566.1985. Neuberger J, Cain O. The Need for Alternatives to Liver Biopsies: Non-Invasive Analytics and Diagnostics. Hepat Med. 2021; 13: 59-69. doi: 10.2147/HMER. Malik S, Das R, Thongtan T, et al. AI in Hepatology: Revolutionizing the Diagnosis and Management of Liver Disease. J Clin Med. 2024; 13: 7833. doi: 10.3390/jcm13247833. Kishanifarahani Z, Ahadi M, Kazeminejad B, et al. Inter-observer Variability in Histomorphological Evaluation of Non-neoplastic Liver Biopsy Tissue and Impact of Clinical Information on Final Diagnosis in Shahid Beheshti University of Medical Sciences Affiliated Hospitals. Iran J Pathol. 2019; 14: 243-247. doi: 10.30699/ijp.2019.99566.1985. Epub 2019 Aug 1. Seven. İ., Bayram, D., Arslan. H, et al. Predicting hepatocellular carcinoma survival with artificial intelligence. Sci Rep.2025; 15, 6226. doi: 10.1038/s41598-025-90884-6. Perincheri S, Levi AW, Celli R. et al. An independent assessment of an artificial intelligence system for prostate cancer detection shows strong diagnostic accuracy. Mod. Pathol. 2021; 34, 1588-1595. doi: 10.1038/s41379-021-00794-x. Calderaro, J., Ghaffari Laleh, N., Zeng. Q, et al. Deep learning-based phenotyping reclassifies combined hepatocellular-cholangiocarcinoma. Nat Commun. 2023; 14: 8290. doi: 10.1038/s41467-023-43749-3. Kalapala R, Rughwani H, Reddy DN. Artificial Intelligence in Hepatology- Ready for the Primetime. J Clin Exp Hepatol. 2023; 13: 149-161. doi: 10.1016/j.jceh.2022.06.009. Lu MY, Williamson DFK, Chen TY, et al. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. 2021; 5: 555-570. doi: 10.1038/s41551-020-00682-w. Chen RJ, Ding T, Lu MY, et al. Towards a general-purpose foundation model for computational pathology. Nat Med. 2024; 30: 850-862. doi: 10.1038/s41591-024-02857-3. Bedossa P, Dargère D, Paradis V. Sampling variability of liver fibrosis in chronic hepatitis C. Hepatology. 2003; 38: 1449-1457. doi: 10.1016/j.hep.2003.09.022. Rousselet MC, Michalak S, Dupre F, et al. Sources of variability in histological scoring of chronic viral hepatitis. Hepatology. 2005; 41: 257-264. doi: 10.1002/hep.20535. Patil A, Salvatori R, Smith L, et al. Artificial intelligence-based reticulin proportionate area - a novel histological outcome predictor in hepatocellular carcinoma. Histopathology. 2023; 83: 512-525. doi: 10.1111/his.15001. Calderaro J, Ghaffari Laleh N, Zeng Q, et al. Deep learning-based phenotyping reclassifies combined hepatocellular-cholangiocarcinoma. Nat Commun. 2023; 14: 8290. doi: 10.1038/s41467-023-43749-3. Pulaski H, Harrison SA, Mehta SS, et al. Clinical validation of an AI-based pathology tool for scoring of metabolic dysfunction-associated steatohepatitis. Nat Med. 2025; 31: 315-322. doi: 10.1038/s41591-024-03301-2. Zeng Q, Klein C, Caruso S, et al. Artificial intelligence-based pathology as a biomarker of sensitivity to atezolizumab-bevacizumab in patients with hepatocellular carcinoma: a multicentre retrospective study. Lancet Oncol. 2023; 24: 1411-1422. doi: 10.1016/S1470-2045(23)00468-0. Sanyal AJ, Loomba R, Anstee QM, et al. Utility of pathologist panels for achieving consensus in NASH histologic scoring in clinical trials: Data from a phase 3 study. Hepatol Commun. 2023; 8: e0325. doi: 10.1097/HC9.0000000000000325. Tables Table 1. Clinical and histopathological characteristics of the training set and the validation sets. Patient characteristics Training Set (n = 673) Validation Set 1 (n = 300) p * value Paired set (pre-treatment) (n = 50) Paired set (post-treatment) (n = 50) p # value Demographic parameters Male sex, n(%） 509 (76%) 231 (77%) 0.644 27 (54%) 32 (64%) --- Age, years 44 (36,51) 43.6 (36.0,51.6) 0.747 39 (33.8, 48.0) 38.0 (32.8, 47.0) --- Serolocical parameters ALT level, U/L 30.7 (21.0, 50.6) 30.9 (21.0, 46.0) 0.926 45.7 (33.0, 84.3) 26.0 (18.9, 40.1) <0.001 AST level, U/L 27.1 (22.0, 38.8) 27.4 (23.0, 36.0) 0.768 35.5 (26.9, 65.5) 22.5 (20.0, 28.2) <0.001 ALB level, U/L 46.0 (43.0, 48.3) 46.2 (43.1, 48.7) 0.756 44.6 (40.8, 47.9) 46.5 (45.0, 48.7) <0.001 TBIL level, μmol/L 14.6 (10.7, 19.5) 14.6 (10.9, 19.2) 0.706 14.7 (12.0, 19.9) 15.1 (10.5, 17.8) 0.486 HBeAg (+), % 31% 28% 0.741 56% 58% 0.705 HBV DNA (+), % 37% 24% 0.015 100% 0% <0.001 PLT ,×10 9 /L 168.0 (129.0, 213.0) 165.0 (124.0, 208.0) 0.247 182.0 (140, 221.5) 190.5 (150.5,224.5) 0.011 LSM, kPa 8.9 (6.4, 12.7) 8.3 (6.4, 10.6) 0.078 9.9 (7.7, 15.5) 6.6 (5.9, 8.9) <0.001 Necroinflammation grade 0.814 <0.001 G0, n (%) 182 (27%) 86 (29%) G0, 3 (6%) 12 (24%) G1, n (%) 281 (42%) 114 (38%) G1, 11 (22%) 23 (46%) G2, n (%) 154 (23%) 68 (23%) G2, 18 (36%) 14 (28%) G3/4, n (%) 56 (8%) 32 (10%) G3/4, 18 (36%) 1 (2%) Fibrosis stage 0.002 <0.001 F0/1, n (%) 279 (41%) 91 (30%) F0/1, 19 (38%) 33 (66%) F2, n (%) 77 (11%) 36 (12%) F2, 5 (10%) 2 (4%) F3, n (%) 218 (32%) 120 (40%) F3, 15 (30%) 12 (24%) F4, n (%) 99 (15%) 53 (18%) F4, 11 (22%) 3 (6%) Note: Data are presented as counts (percentages) or medians (with interquartile ranges, IQR) for continuous variables with skewed distributions. The p * value indicates differences between the training set and Validation Set 1, analyzed using the Mann-Whitney U test. The p # value indicates differences between the paired sets before and after treatment, analyzed using the Wilcoxon Signed-Rank Test. Table 2. Performance of the BJ-HepaGS model in histopathological grading and staging in independent validation set 1 and paired sets. Accuracy Sensitivity Specificity AUROC (95%CI) Validation set 1 Inflammation ≥G1 0.890 0.957 0.720 0.91 (0.87-0.95) ≥G2 0.876 0.800 0.915 0.95 (0.92-0.97) ≥G3 0.946 0.718 0.973 0.98 (0.97-0.99) Fibrosis ≥F2 0.850 0.895 0.748 0.89 (0.84-0.93) ≥F3 0.810 0.931 0.646 0.87 (0.83-0.91) F4 0.847 0.340 0.956 0.82 (0.76-0.88) Validation paired set Inflammation ≥G1 0.910 0.953 0.667 0.92 (0.85-0.97) ≥G2 0.900 0.902 0.898 0.98 (0.96-0.99) ≥G3 0.940 0.842 0.963 0.96 (0.92-0.99) Fibrosis ≥F2 0.780 0.875 0.692 0.90 (0.84-0.95) ≥F3 0.810 0.927 0.729 0.92 (0.87-0.97) F4 0.870 0.571 0.919 0.91 (0.82-0.98) Table 3. Concordance of the BJ-HepaGS model and expert consensus Histology features ICC of AI model vs. Consensus ICC of Senior experts scores vs. Consensus ICC of Junior experts scores vs. Consensus Validation set 1 Inflammation 0.824 (0.784-0.857) 0.961 (0.952-0.969) 0.763 (0.711-0.807) ≥G1 0.721 (0.662-0.771) 0.984 (0.980-0.987) 0.477 (0.385-0.560) ≥G2 0.721 (0.661-0.771) 0.917 (0.897-0.933) 0.755 (0.702-0.800) ≥G3 0.712 (0.652-0.764) 0.930 (0.913-0.944) 0.756 (0.703-0.801) Fibrosis 0.681 (0.615-0.737) 0.931 (0.915-0.945) 0.912 (0.891-0.930) ≥F2 0.644 (0.573-0.706) 0.914 (0.893-0.930) 0.905 (0.883-0.924) ≥F3 0.613 (0.537-0.679) 0.888 (0.861-0.909) 0.863 (0.831-0.889) F4 0.369 (0.267-0.463) 0.829 (0.790-0.862) 0.804 (0.760-0.840) Validation paired set Inflammation 0.865 (0.805-0.907) 0.942 (0.915-0.960) 0.831 (0.758-0.883) ≥G1 0.637 (0.505-0.741) 0.962 (0.944-0.974) 0.618 (0.480-0.726) ≥G2 0.800 (0.716-0.861) 0.841 (0.772-0.890) 0.574 (0.427-0.692) ≥G3 0.805 (0.723-0.865) 0.934 (0.903-0.955) 0.754 (0.655-0.827) Fibrosis 0.682 (0.562-0.775) 0.930 (0.897-0.952) 0.910 (0.870-0.939) ≥F2 0.574 (0.427-0.692) 0.920 (0.883-0.945) 0.840 (0.771-0.889) ≥F3 0.647 (0.517-0.748) 0.938 (0.909-0.958) 0.896 (0.850-0.929) F4 0.476 (0.309-0.614) 0.812 (0.733-0.870) 0.722 (0.614-0.804) Table 4. The BJ-HepaGS model improved the consistency of scores between senior and junior experts. Histology features ICC of Senior vs. Junior experts scores ICC of Senior vs. Junior experts scores refined by AI model Validation set 1 Inflammation 0.730 (0.673-0.779) 0.818 (0.777-0.852) ≥G1 0.476 (0.383-0.559) 0.460 (0.366-0.545) ≥G2 0.684 (0.619-0.740) 0.835 (0.798-0.867) ≥G3 0.659 (0.590-0.719) 0.822 (0.782-0.856) Fibrosis 0.852 (0.817-0.880) 0.904 (0.881-0.922) ≥F2 0.819 (0.778-0.853) 0.860 (0.827-0.886) ≥F3 0.766 (0.715-0.809) 0.889 (0.863-0.911) F4 0.718 (0.659-0.769) 0.793 (0.747-0.831) Additional Declarations There is NO Competing Interest. Supplementary Files Supplementarymaterials.docx 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-7440962","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":509496814,"identity":"4b1210b3-b98a-4ec4-b797-03b951f2fd47","order_by":0,"name":"Xinyan Zhao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8ElEQVRIiWNgGAWjYDACCQTJ+AAqZkC0FmaYUqK0gAEbjI1fi/zs5ocPGNss8gyOnz1W8aNmW2IDe/M2CYaaOzi1MM45ZmzAcEai2OBMXtrNnmO3Ext4jpVJMBx7hlMLs0SCmQRDhUTihgM5ZrcZG4BaJHLMJBgbDuPUwiaR/k2CwQCo5fwbs2KwFvk3+LXwgMwE23Ijx4wZYgsPfi0SEjnFIL8kzrzxxlgS6BfjNp60YouEY7i1yM9I3wgMsbrEvvM5hh9+1NyW7Wc/vPHGhxrcWsBB8AdIKByA+Q5EJODVALOugRhVo2AUjIJRMCIBAJrWUo9gjkCIAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0002-8016-4368","institution":"Beijing Friendship Hospital, Capital Medical University","correspondingAuthor":true,"prefix":"","firstName":"Xinyan","middleName":"","lastName":"Zhao","suffix":""},{"id":509496815,"identity":"bd539dc0-0ac8-428c-8415-f149ec21ea9f","order_by":1,"name":"Jialing Zhou","email":"","orcid":"","institution":"Beijing Friendship Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Jialing","middleName":"","lastName":"Zhou","suffix":""},{"id":509496816,"identity":"56d796b3-41ec-4534-a5bc-c0d60cf9db4b","order_by":2,"name":"Ying Zhang","email":"","orcid":"","institution":"Tianjin Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Zhang","suffix":""},{"id":509496817,"identity":"7e4a3348-8bb3-4b5d-9e60-1bdece65f941","order_by":3,"name":"Yongyin Li","email":"","orcid":"https://orcid.org/0000-0001-6303-7642","institution":"Nanfang Hospital, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yongyin","middleName":"","lastName":"Li","suffix":""},{"id":509496818,"identity":"fe067a49-a2a4-47d3-9914-be5e57a0482d","order_by":4,"name":"Lin Wang","email":"","orcid":"","institution":"Beijing Friendship Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Wang","suffix":""},{"id":509496819,"identity":"151dd55b-0e9a-4d1e-b9be-a14e81dd4a2a","order_by":5,"name":"Guoxin Teng","email":"","orcid":"","institution":"Shandong University","correspondingAuthor":false,"prefix":"","firstName":"Guoxin","middleName":"","lastName":"Teng","suffix":""},{"id":509496820,"identity":"be138c99-8e69-467b-bc79-c180206e868a","order_by":6,"name":"Xiaoli Li","email":"","orcid":"","institution":"Linyi People’Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiaoli","middleName":"","lastName":"Li","suffix":""},{"id":509496821,"identity":"ff0c4ecb-cf6b-4361-a12d-a43a3c72d3a2","order_by":7,"name":"Song Yang","email":"","orcid":"","institution":"Department of Hepatology, Beijing Ditan Hospital of Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Song","middleName":"","lastName":"Yang","suffix":""},{"id":509496822,"identity":"5dc9f9a8-0585-496b-90ac-d6bd4e80bdca","order_by":8,"name":"Hui Liu","email":"","orcid":"","institution":"Department of Pathology, Beijing You’an Hospital, Affiliated with Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"Liu","suffix":""},{"id":509496823,"identity":"eeee3f73-3109-4d49-a908-2adfc21de6b6","order_by":9,"name":"Fangzhi Li","email":"","orcid":"","institution":"Tianjin Medical University","correspondingAuthor":false,"prefix":"","firstName":"Fangzhi","middleName":"","lastName":"Li","suffix":""},{"id":509496824,"identity":"62702971-efc3-4cdb-abb5-801cf142ec58","order_by":10,"name":"Xiaojuan Ou","email":"","orcid":"","institution":"Beijing Friendship Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiaojuan","middleName":"","lastName":"Ou","suffix":""},{"id":509496825,"identity":"f434c790-e184-4f5c-b409-317ecfadbd1c","order_by":11,"name":"Hong You","email":"","orcid":"https://orcid.org/0000-0001-9409-1158","institution":"Beijing Friendship Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Hong","middleName":"","lastName":"You","suffix":""},{"id":509496826,"identity":"43249916-4b00-4673-8878-9cddbc2f5008","order_by":12,"name":"Ji-Dong Jia","email":"","orcid":"","institution":"Liver Research Center, Beijing Friendship Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ji-Dong","middleName":"","lastName":"Jia","suffix":""},{"id":509496827,"identity":"702d499e-90e7-4bd8-8bce-09772b94256a","order_by":13,"name":"Chunhong Hu","email":"","orcid":"https://orcid.org/0000-0003-4061-1183","institution":"Tianjin Medical University","correspondingAuthor":false,"prefix":"","firstName":"Chunhong","middleName":"","lastName":"Hu","suffix":""}],"badges":[],"createdAt":"2025-08-23 11:30:09","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7440962/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7440962/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":91565325,"identity":"f2cfb9e2-2168-4b03-884a-aa324aaaeb8e","added_by":"auto","created_at":"2025-09-17 19:23:01","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":555498,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe workflow of grading inflammation and staging fibrosis both by HE staining using H\u0026amp;E-stained WSIs and a vision transformer. \u003c/strong\u003eAll WSIs were reviewed by pathologists, divided into training and validation sets. These patches were encoded with a pre-trained vision transformer and aggregated via a gated-attention mechanism within the CLAM framework to develop the models. Model performance was evaluated using ROC curves, confusion matrices and a paired comparison. \u003cem\u003eH\u0026amp;E, Hematoxylin and Eosin; WSIs, Whole Slide Images; whole slide images; CLAM, clustering-constrained attention multiple instance learning model; ROC, receiver operating characteristic curve;\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-7440962/v1/fa2d4a9dfb085d4fad7bb70c.png"},{"id":91565890,"identity":"0ff46baf-0bef-4170-97af-04e2e45d5be3","added_by":"auto","created_at":"2025-09-17 19:31:01","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":42853,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFlowchart of the WSIs screening and sets division. \u003c/strong\u003eA total of 3,219 available liver biopsy slides (each including H\u0026amp;E, Masson's Trichrome, and Reticulin staining) from 1023 patients across four centers were divided into a training set, a validation set and a paired liver biopsies set. CHB, Chronic Hepatitis B; \u003cem\u003eWSIs, Whole Slide Images; H\u0026amp;E, Hematoxylin and Eosin. WSIs, whole slide images;\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-7440962/v1/bb0caf26f4a645e5070b777b.png"},{"id":91565891,"identity":"3d90ccc8-814b-4908-a981-593fafc28a55","added_by":"auto","created_at":"2025-09-17 19:31:01","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1620871,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualization of high-attention patch selection and heatmaps for different stages of inflammation and fibrosis in the BJ-HepaGS model. \u003c/strong\u003eThis figure illustrated the model’s interpretability at the patch level through its attention-based mechanism. (A) An original H\u0026amp;E-stained WSI was divided into patches. (B) Representative high-attention patches for inflammation grades (G1–G4) and fibrosis stages (F1–F4). (C) Two-dimensional PCA projection of patch embeddings, with color intensity indicating attention weights. (D) Attention heatmaps highlighted regions associated with varying stages of inflammation and fibrosis as identified by the BJ-HepaGS model.\u003cem\u003e H\u0026amp;E, Hematoxylin and Eosin; WSIs, Whole Slide Images; PCA, principal component analysis;\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-7440962/v1/b1a2ca4ebfdf8b2942aac31d.png"},{"id":91565324,"identity":"55939a65-da58-42e9-8686-d716c28c954e","added_by":"auto","created_at":"2025-09-17 19:23:01","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":81539,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance evaluation of the BJ-HepaGS model across ROC analysis, expert interpretations refined by AI model, and histopathological changes. \u003c/strong\u003e(A) ROC curves for inflammation grading and fibrosis staging in two independent validation cohorts. (B) Influence of AI model on experts interpretations. (C) Assessment of the inflammation improvement and fibrosis regression before and after treatment.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-7440962/v1/61edde7396813e32000982a7.png"},{"id":91565322,"identity":"adc3de94-7b17-4ac8-b384-190321ca9e6b","added_by":"auto","created_at":"2025-09-17 19:23:01","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":127844,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConfusion matrices of the BJ-HepaGS model’s performance in inflammation grading and fibrosis staging across the validation sets.\u003c/strong\u003e These matrices illustrate three binary classification schemes for both inflammation grade and fibrosis stage in validation set 1 and the paired validation set. Color intensity reflects the number of predicted samples, with darker shades indicating higher counts.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-7440962/v1/90bc9aca932299d1c383a9c7.png"},{"id":93509121,"identity":"fe54138c-e730-4abb-bd4f-63c05589f6e1","added_by":"auto","created_at":"2025-10-14 15:18:52","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4114704,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7440962/v1/061efd80-980e-496f-8fee-aa7a442daed2.pdf"},{"id":91564770,"identity":"06336a7b-7c03-491e-aed4-4edd30838d5d","added_by":"auto","created_at":"2025-09-17 19:15:01","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":172121,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-7440962/v1/7cee9d9305f75a7ce1245163.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Artificial Intelligence Model for Accurate Grading and Staging of Chronic Hepatitis B: Development and Validation","fulltext":[{"header":"Introduction","content":"\u003cp\u003eChronic hepatitis B (CHB) remains a major global health challenge, with approximately 254\u0026nbsp;million people living with chronic HBV infection as of 2022.\u003csup\u003e[1]\u003c/sup\u003e CHB is responsible for 820,000 to 1.4\u0026nbsp;million annual deaths worldwide, primarily due to complications such as cirrhosis and hepatocellular carcinoma (HCC).\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e The burden is particularly pronounced in the Asian Pacific, where prevalence rates are highest, and in low- and middle-income countries, where the economic effect is substantial.\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e To address this issue, the World Health Organization (WHO) has set ambitious targets to eliminate viral hepatitis as a public health threat by 2030; WHO aims to reduce new HBV infections by 90% and HBV-related deaths by 65% compared to 2015 levels.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eAccurate assessment of inflammatory activity (grading) and liver fibrosis (staging) is critical for the effective management of CHB.\u003csup\u003e\u003cspan additionalcitationids=\"CR8 CR9 CR10\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e Liver biopsy remains the gold standard for grading and staging because it provides detailed histopathological information essential for clinical decision-making.\u003csup\u003e\u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e Reliable histological evaluation also underpins the development of noninvasive models for predicting inflammatory grading and fibrosis staging. However, the clinical utility of liver biopsy is limited by significant inter- and intra-observer variability in histopathological interpretation, which undermines consistent disease assessment and management.\u003csup\u003e\u003cspan additionalcitationids=\"CR16 CR17\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e These limitations highlight the urgent need for more objective and reproducible diagnostic approaches to improve patient care.\u003c/p\u003e\u003cp\u003eWhile artificial intelligence (AI) has shown promise in enhancing diagnostic accuracy and standardization in hepatology, its application in liver histopathology has primarily focused on HCC diagnosis.\u003csup\u003e\u003cspan additionalcitationids=\"CR20 CR21\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e The majority of AI-driven studies in this field have concentrated on improving the detection and characterization of HCC; to our knowledge, few published AI studies have focused on non-tumor liver diseases such as CHB. This paucity of research underscores a critical unmet need in the field.\u003c/p\u003e\u003cp\u003eIn this study, we aimed to develop and validate a novel AI-based model to enhance the accuracy and reliability of CHB grading and staging, ultimately contributing to better clinical care for patients with CHB.\u003c/p\u003e"},{"header":"Patients and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003ePatient Population and Study Design\u003c/h2\u003e\u003cp\u003eThis multi-center, retrospective study enrolled patients with CHB from four Chinese hospitals from January 2013 to December 2024. The study was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (2024-P2-042-01). Written informed consent was waived. All procedures were conducted in accordance with the Declaration of Helsinki and relevant local regulations.\u003c/p\u003e\u003cp\u003eA total of 3,219 liver biopsy slides, each including hematoxylin and eosin (H\u0026amp;E), Masson\u0026rsquo;s trichrome, and reticulin staining, were included (1,023 patients). The slides represented patients with diverse inflammation grades and fibrosis stages, before or after treatment. The inclusion criteria were as follows: (a) age\u0026thinsp;\u0026ge;\u0026thinsp;18 years, regardless of gender; (b) CHB diagnosis (hepatitis B surface antigen positive for \u0026gt;\u0026thinsp;6 months); (c) pathological liver biopsy indicated a length\u0026thinsp;\u0026gt;\u0026thinsp;1.0 cm and/or more than six portal tracts. The exclusion criteria were as follows: (a) co-occurrence of any other liver diseases, including autoimmune hepatitis, alcoholic liver disease, or HCC at the time of liver biopsy; (b) co-infection with any other viral hepatitis, such as hepatitis C or D; and (c) poor staining quality of the liver biopsy samples.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eSample Selection and Interpretation\u003c/h3\u003e\n\u003cp\u003eLiver biopsy was performed in the right lobe of the liver using a 16- or 18-gauge needle. The specimens were stained at the different centers using H\u0026amp;E, Masson\u0026rsquo;s trichrome, and reticulum staining to assess necroinflammatory activity and fibrosis. Liver biopsy slides were uniformly scanned using a consistent device (3DHISTECH, Budapest, Hungary) and digitally processed for AI analysis.\u003c/p\u003e\n\u003ch3\u003eInterpretation Methods\u003c/h3\u003e\n\u003cp\u003eFour experts\u0026mdash;three pathologists and one hepatologist with specialized pathological training\u0026mdash;were divided into two groups (Senior Group A: L. Wang and X.Y. Zhao; Junior Group B: X.L. Li and G.X. Teng). Each group independently interpreted the whole-slide images (WSIs) for inflammation grading and fibrosis staging according to the METAVIR system. Discrepancies were resolved through re-evaluation to achieve consensus. The consensus interpretation served as the ground truth for training the AI model.\u003c/p\u003e\n\u003ch3\u003eClinical Data Collection\u003c/h3\u003e\n\u003cp\u003eClinical data, collected at the time of pathological examination, included demographic parameters (age and gender) and the following laboratory parameters: platelet count (PLT), alanine aminotransferase (ALT), aspartate aminotransferase (AST), serum albumin (ALB), total bilirubin (TBIL), HBV DNA, and HBeAg status.\u003c/p\u003e\n\u003ch3\u003eDevelopment of the AI Model\u003c/h3\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eAI-based Analysis Method\u003c/h2\u003e\u003cp\u003eWe employed the clustering-constrained attention multiple instance learning model (CLAM),\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e a robust AI approach for WSI analysis. The CLAM model consists of four key components (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e):\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eFeature Extraction\u003c/span\u003e: A pretrained UNI feature extractor encodes image patches into representative features,\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e leveraging its ability to learn diverse medical image patterns.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eGated Attention Layer\u003c/span\u003e: An attention layer assigns attention scores to image patches, prioritizing regions critical for classification.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eClustering Constraint\u003c/span\u003e: Pseudo-labels for high- and low-attention patches are generated during training, enhancing supervisory signals and feature representation.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eFully Connected Layer\u003c/span\u003e: Global feature representations are mapped to class labels, completing the classification task.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eDataset Allocation\u003c/h3\u003e\n\u003cp\u003eA total of 673 H\u0026amp;E WSIs were used for training. Two external datasets were used to validate the model: one dataset of 300 H\u0026amp;E WSIs from four centers; and a paired liver biopsy dataset of 100 H\u0026amp;E WSIs. These datasets were employed to train and validate the model, ensuring accurate grading and staging of inflammation and fibrosis based on H\u0026amp;E staining.\u003c/p\u003e\n\u003ch3\u003eTraining Process\u003c/h3\u003e\n\u003cp\u003eWSIs were divided into non-overlapping 512\u0026times;512-pixel patches, separating tissue regions from background pixels. The UNI extractor was applied to extract features, followed by attention score computation via the gated attention layer. Pseudo-labels were generated through clustering constraints, and features were aggregated using attention-weighted representations for the final task. The model incorporates five attention branches to assess tissue patch significance, enhancing interpretability. Performance was evaluated based on the area under the receiver operating characteristic curve (AUROC) and accuracy. Training was conducted using an NVIDIA DGX A100 GPU with an 80-GB memory capacity.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eWe used various statistical methods to analyze the data and assess the AI model performance. Continuous variables (e.g., age and biochemical markers) were summarized by median, and interquartile range, while categorical variables (e.g., gender) were described by frequencies and percentages. Categorical comparisons were made using chi-square or Fisher\u0026rsquo;s exact tests, with two-sided \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05 indicating significance. These analyses were conducted using SPSS (version 26.0) and GraphPad Prism (version 6.02).\u003c/p\u003e\u003cp\u003eWhen comparing AI model performance with expert scores, paired t-tests were applied for continuous variables and chi-square tests for categorical variables. ROC curves were generated using MedCale software (version 15.8) to evaluate the diagnostic accuracy of the AI model for inflammation grading and fibrosis staging. Model generalizability was tested on an independent validation set using metrics of accuracy and specificity. A paired dataset was also used to assess the ability of the model to evaluate dynamic changes of grading and staging. The AI model was developed and optimized using Python (version 3.10.4).\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eA total of 1,217 patients were screened for this study. After excluding 194 patients due to unqualified histological samples and missing serological results, 1023 patients, including 3,219 WSIs, were included in the final analysis (\u003cstrong\u003eFig. 2\u003c/strong\u003e). The dataset was divided into three sets: 673 WSIs (627,703 patches) for the training set, 300 WSIs (331,037 patches) for the validation set, and 100 WSIs (52,271 patches) for the paired set.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eClinical and Histopathological Characteristics of the Training and Validation Sets\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 673 WSIs were included in the training dataset to develop the BJ-HepaGS model. The ratio of patients before treatment to those after treatment was approximately 2:1. The cohort was predominantly male (76%). Among the patients, 37% and 31% exhibited HBV DNA and were HBeAg positive, respectively. A total of 463 patients (69%) had no or minimal inflammation and were graded as G0 or G1. Moderate inflammation (G2) was observed in 23% of patients, while severe inflammatory activity (G3/4) was presented in the smallest proportion of patients (8%). Regarding fibrosis staging, 279 (41%) patients were staged at F0 or F1. A total of 77 (11%) patients were staged at F2. Advanced fibrosis accounted for a larger proportion, with 218 (32%) patients at F3 and 99 (15%) at F4.\u003c/p\u003e\n\u003cp\u003eThe ratio of patients before treatment to patients after treatment in the independent validation set was approximately 1:5. No significant differences were observed between the training set and validation set 1 in terms of inflammation grade, ALT, AST, ALB, TBIL, PLT, or HBeAg positivity (\u003cem\u003ep\u003c/em\u003e \u0026gt; 0.05). However, a significant difference was observed in the distribution of fibrosis stages (\u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05), which might be attributed to the higher proportion of post-treatment cases in this validation set.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe paired validation set included 50 cases (100 H\u0026amp;E WSIs), with both pre-treatment and post-treatment samples available. Significant improvements in both inflammation and fibrosis were observed after treatment. Prior to treatment, the distribution of grades was as follows: 14 patients (28%) graded G0 or G1, 18 (36%) G2, and 18 (36%) G3/4. In terms of fibrosis stages, 19 patients (38%) were staged at F0 or F1, 5 (10%) at F2 , 15 (30%) at F3, and 11 (22%) at F4. The baseline characteristics of all patients were summarized in \u003cstrong\u003eTable 1\u003c/strong\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eVisualization of High-attention Patch Selection and Heatmaps Across Inflammation Grades and Fibrosis stages in the BJ-HepaGS Model\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe overall workflow of the BJ-HepaGS model for analyzing H\u0026amp;E-stained liver biopsy slides was illustrated in\u003cstrong\u003e\u0026nbsp;Fig. 3A\u003c/strong\u003e. The WSIs were first divided into fixed-size patches, and features were extracted from each individual patch. The model identified high-attention patches across various stages of inflammation (G1\u0026ndash;G4) and fibrosis (F1\u0026ndash;F4), as illustrated in \u003cstrong\u003eFig. 3B\u003c/strong\u003e. To further assess the distribution of learned features, we conducted principal component analysis (PCA) on the patch-level embeddings derived from a representative WSI. As illustrated in the bottom panel, patches with higher attention scores (indicated by darker colors) tended to cluster together in the reduced feature space (\u003cstrong\u003eFig. 3C\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003eThe attention-based heatmap visualizations generated by the model over the WSIs are shown in \u003cstrong\u003eFig. 3D\u003c/strong\u003e. Color overlays on each WSI indicate the attention strength assigned by the model, with red denoting high attention and blue indicating low attention. The magnified views highlight a strong correspondence between high-attention regions and histologically relevant features, suggesting that the model effectively localized tissue patterns associated with inflammation and fibrosis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003ePerformance of the BJ-HepaGS Model in Grading and Staging: An Independent Validation Set\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the validation set of 300 H\u0026amp;E WSIs, we assessed the performance of the BJ-HepaGS model by evaluating its diagnostic performance and comparing its outputs to the consensus of expert pathologists (\u003cstrong\u003eFig. 4\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eGrading\u003c/u\u003e:\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThe BJ-HepaGS model performed well in grading. The model accurately identified inflammation of grades \u0026ge; G1 (AUROC = 0.91; 95% CI: 0.87\u0026ndash;0.95), \u0026ge; G2 (AUROC = 0.95; 95% CI: 0.92\u0026ndash;0.97), and \u0026ge; G3 (AUROC = 0.98; 95% CI: 0.97\u0026ndash;0.99). The model also achieved high accuracy for assessing all grades: \u0026ge; G1 (89.0%), \u0026ge; G2 (87.6%), and \u0026ge; G3 (94.6%) \u003cstrong\u003e(Table 2)\u003c/strong\u003e. The confusion matrices also indicated that the model correctly classified grades \u0026ge; G3 \u003cstrong\u003e(Fig. 5).\u0026nbsp;\u003c/strong\u003e .\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eStaging\u003c/u\u003e:\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThe BJ-HepaGS model demonstrated good performance in staging. In particular, the model excelled in identifying advanced stages: \u0026ge; F2 (AUROC = 0.89; 95% CI: 0.84\u0026ndash;0.93), \u0026ge; F3 (AUROC = 0.87; 95% CI: 0.83\u0026ndash;0.91), and F4 (AUROC = 0.82; 95% CI: 0.76\u0026ndash;0.88). The model also showed good accuracy for assessing all stages: \u0026ge; F2 (85.0), \u0026ge; F3 (81.0%), and F4 (84.7%). \u003cstrong\u003e(Table 2)\u003c/strong\u003e The confusion matrices also demonstrated that the model correctly classified stages \u0026ge; F2 (\u003cstrong\u003eFig. 5\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eConcordance of the BJ-HepaGS Model with Expert Consensus in Grading and Staging\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the validation set of 300 H\u0026amp;E WSIs and the paired validation set of 100 H\u0026amp;E WSIs, we assessed the concordance of the BJ-HepaGS model and expert consensus of ICC (\u003cstrong\u003eTable 3\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cu\u003eGrading\u003c/u\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThe model demonstrated strong agreement with the expert consensus for inflammation grading, achieving an ICC of 0.824 (95% CI: 0.784\u0026ndash;0.857). The performance of the model remained consistent across different grades: G \u0026ge; 1 (ICC = 0.721; 95% CI: 0.662\u0026ndash;0.771), G \u0026ge; 2 (ICC = 0.721; 95% CI: 0.661\u0026ndash;0.771), and G \u0026ge; 3 (ICC = 0.712; 95% CI: 0.652\u0026ndash;0.764). In the analysis of 50 paired samples, the BJ-HepaGS model achieved an ICC of 0.865 (95% CI: 0.805\u0026ndash;0.907). The AI model assessments of inflammation were closely aligned with the consensus of human experts (\u003cstrong\u003eTable 2\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cu\u003eStaging\u003c/u\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThe model also achieved a moderate ICC of 0.681 (95% CI: 0.615\u0026ndash;0.737) for staging. It showed consistency for fibrosis staging: F \u0026ge; 2 (ICC = 0.644; 95% CI: 0.573\u0026ndash;0.706), F \u0026ge; 3 (ICC = 0.613; 95% CI: 0.537\u0026ndash;0.679) and F4 (ICC = 0.369; 95% CI: 0.267\u0026ndash;0.463). In the 50 paired samples, the model also demonstrated moderate consistency for fibrosis staging, with an ICC of 0.682 (95% CI: 0.562\u0026ndash;0.775).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAssessment of inflammation improvement and fibrosis regression in paired pre- and post-treatment samples\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInflammation improvement was defined as a decrease of 1 or more in the inflammation grade; stable inflammation was defined as no change; and worsening inflammation was defined as an increase of 1 or more in grade. According to the BJ-HepaGS model, 33 cases (66%) demonstrated improvement, 13 cases (26%) remained stable, and 4 cases (8%) exhibited worsening. Based on expert consensus, 34 cases (68%) showed improvement, 11 cases (22%) remained stable, and 5 cases (10%) worsened. As for the change in inflammation, the distribution of outcomes between the AI model and expert consensus was consistent, with no statistically significant difference (\u003cem\u003ep\u003c/em\u003e = 0.885; \u003cstrong\u003eFig. 4\u003c/strong\u003e and \u003cstrong\u003eSupplementary Table 1\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003eFibrosis regression was defined as a decrease of 1 or more in the fibrosis stage; stability was defined as no change; and progression was defined as an increase of 1 or more. The\u003cem\u003e\u0026nbsp;\u003c/em\u003eBJ-HepaGS model identified fibrosis regression in 22 cases (44%), stability in 25 cases (50%), and progression in 3 cases (6%). Expert consensus found regression in 23 cases (46%), stability in 23 cases (46%), and progression in 4 cases (8%). As for the change in fibrosis, the AI model and expert consensus was consistent with no statistically significant difference (\u003cem\u003ep\u003c/em\u003e = 0.388; \u003cstrong\u003eFig. 4\u003c/strong\u003e and \u003cstrong\u003eSupplementary Table 1\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003eThese results demonstrate that the BJ-HepaGS model is capable of interpreting dynamic changes in CHB, supporting its potential as a reliable tool for evaluating treatment response and monitoring disease progression.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eEnhanced Concordance with Expert Interpretations After Iterative Refinement\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the validation set of 300 cases, discrepancies between the BJ-HepaGS model and expert consensus were observed in approximately 20% of inflammation assessments and 40% of fibrosis assessments. Notably, the AI model provided substantial assistance to both the senior and junior expert groups during re-evaluation (\u003cstrong\u003eFig. 4\u003c/strong\u003e and \u003cstrong\u003eSupplementary Table 2\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cu\u003eGrading\u003c/u\u003e\u003c/strong\u003e\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003eAfter reviewing the suggestions from the AI model, 44% (26/59) of senior experts and 46% (32/70) of junior experts revised their grading scores. Among those who changed their scores, most of the senior experts (88%, 23/26) and junior experts (88%, 28/32) aligned their revised scores with the AI model; only a small minority adjusted their scores to alternative scores.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cu\u003eStaging\u003c/u\u003e\u003c/strong\u003e\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003eAfter reviewing the suggestions from the AI model, 29% (34/117) of senior experts and 31% (40/128) of junior experts revised their scores. Among those who changed their scores, most of the senior experts (76%, 26/34) and an even higher proportion of the junior experts (80%, 32/40) aligned their revised scores with the AI model; only a small minority adjusted their scores to alternative scores.\u003c/p\u003e\n\u003cp\u003eIterative AI assistance improved agreement between senior and junior expert scores, resulting in a model closely matching expert grading. For inflammation, the ICC increased from 0.730 (95% CI: 0.673-0.779) to 0.857 (95% CI: 0.824\u0026ndash;0.884). For fibrosis, the ICC increased from 0.852 [95% CI: 0.817\u0026ndash;0.880) to 0.906 [95% CI: 0.883\u0026ndash;0.924]. Concordance remained high in cirrhosis (ICC 0.822, 95% CI: 0.782\u0026ndash;0.856; Table 3). These findings demonstrate that the AI model substantially improved the precision and reliability of expert interpretations, enabling more accurate liver biopsy assessments (Table 4).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe BJ-HepaGS model, based on only H\u0026amp;E-stained WSIs, demonstrated expert-level accuracy in grading and staging CHB. The model achieved an AUROC of 0.91 for necroinflammation grading and 0.81 for fibrosis staging, indicating high performance relative to expert consensus. Furthermore, in the paired validation set, the model exhibited significant correlations with expert assessments of changes in necroinflammation grade (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.855) and fibrosis stage (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.388). In approximately one-third of the cases where discrepancies existed between the AI model and the scores of different groups of liver pathology experts, the implementation of the BJ-HepaGS model facilitated modifications in the expert scores, underscoring its capacity to reduce both inter-observer and intra-observer variability.\u003c/p\u003e\u003cp\u003eLiver biopsy remains the gold standard for assessing CHB. Liver biopsy serves as the foundation for noninvasive model development, treatment decision-making, and outcome prediction. However, histopathological assessment is subject to substantial inter-observer variability.\u003csup\u003e\u003cspan additionalcitationids=\"CR26\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e As the first AI model specifically designed to provide objective and reproducible assessments, the BJ-HepaGS model represents significant progress toward addressing this limitation. Notably, the model achieves near-expert consistency, with an ICC of 0.824 for inflammation grading and 0.68 for fibrosis staging, demonstrating its potential to reduce variability in CHB assessment.\u003c/p\u003e\u003cp\u003eCurrently, machine learning models are primarily implemented in various pathology applications, including cancer detection, tumor origin identification, and metabolic dysfunction-associated steatohepatitis (MASH).\u003csup\u003e\u003cspan additionalcitationids=\"CR29\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e Sanyal et al. recently demonstrated the strong performance of AIM-MASH (AI-based measurement of MASH) in histological assessments, achieving kappa values of 0.67 (95% CI: 0.64\u0026ndash;0.71) for lobular inflammation and 0.62 (95% CI: 0.58\u0026ndash;0.65) for fibrosis.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e Notably, the performance of the AIM-MASH tool is comparable to that of our BJ-HepaGS model, highlighting the potential of deep learning in pathological assessments. By leveraging advanced deep learning techniques, these AI-based models can enhance the precision and consistency of histological evaluations while minimizing the subjectivity and variability inherent in manual scoring systems.\u003c/p\u003e\u003cp\u003eThe advanced design of our model enables it to focus on diagnostically relevant regions within H\u0026amp;E-stained WSIs, thereby enhancing both interpretability and accuracy. The AIM-MASH system analyzes WSIs using both H\u0026amp;E and Masson\u0026rsquo;s trichrome staining and employs convolutional neural networks to generate color-coded overlays for histological feature segmentation.\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e In contrast, our model operates on routine H\u0026amp;E-stained slides alone; thus, by eliminating the need for special stains like Sirius Red, our model improves clinical practicality and scalability. These technical advantages position the BJ-HepaGS model as a robust and deployable solution for AI-assisted liver pathology.\u003c/p\u003e\u003cp\u003eA major strength of this study is the use of a large, diverse, and multi-center dataset, which enhanced the generalizability of the BJ-HepaGS model. Standardized histopathological criteria were applied throughout, ensuring consistency and reliability in both model training and evaluation. The performance of the model was rigorously benchmarked against expert consensus, demonstrating robust accuracy in cross-sectional grading and staging as well as in the longitudinal assessment of histological changes, an area where many existing models fall short due to limited validation or lack of dynamic assessment.\u003c/p\u003e\u003cp\u003eThis study has several limitations. While the model performed well in inflammation grading, its accuracy for fibrosis staging was lower, likely due to the challenge of detecting subtle changes on H\u0026amp;E stains alone. Expanding the training set with more severe fibrosis cases and incorporating fibrosis-specific stains, such as Masson\u0026rsquo;s trichrome, may improve performance. Additionally, since most patients were from Chinese hospitals, results may not be generalizable to other populations.\u003c/p\u003e\u003cp\u003e Successful integration of BJ-HepaGS into clinical practice will require ensuring compatibility with digital pathology systems, effective user training, and regular model updates as guidelines evolve. Prospective, real-world studies are needed to validate its impact on clinical decision-making and patient outcomes. Future research should also explore combining AI-based histology with clinical and laboratory data to provide comprehensive decision support for CHB management.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn summary, the BJ-HepaGS model represents a significant advancement in AI-assisted liver pathology. The model offers expert-level accuracy, reproducibility, and the ability to monitor dynamic histological changes. By addressing key challenges in the assessment of liver biopsy slides, this model has the potential to standardize care, improve diagnostic confidence, and ultimately enhance outcomes for patients with CHB.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eCHB, chronic hepatitis B; HBV, hepatitis B virus; AI, artificial intelligence; H\u0026amp;E, hematoxylin and eosin; WSIs, whole-slide images; CLAM, clustering-constrained attention multiple instance learning model; ALB, albumin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; TBIL, total bilirubin; AUROC, area under the receiver operating characteristic curve; HBeAg, Hepatitis B e antigen positive; LSM, liver stiffness measurement; PLT, platelet count; ROC, receiver operating characteristic curve; ULN, upper limit of normal; SD, standard deviation; IQR, interquartile range; CI, confidence interval; ICC, Intraclass Correlation Coefficient.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor\u003c/strong\u003e\u003cstrong\u003es\u0026rsquo; c\u003c/strong\u003e\u003cstrong\u003eontribution:\u0026nbsp;\u003c/strong\u003eJ.L. Zhou, Y Zhang and Y.Y. Li contribute equally to the study. X.Y. Zhao designed the study, with assistance from J.L. Zhou. Y.Y. Li, S. Yang, H. Liu collected the data. X.Y. Zhao, L Wang, G.X. Teng and X.L. Li interpret the inflammation and fibrosis. C.H. Hu and Y. Zhang trained the model. J.L. Zhou prepared the graphs and drafted the manuscript. X.Y. Zhao supervised the work, reviewed the manuscript, and provided critical insights during the editing process. All co-authors reviewed, revised, and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Transparency Statement:\u0026nbsp;\u003c/strong\u003eData will be made available on request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eC\u003c/strong\u003e\u003cstrong\u003eonflict of interest statements\u003c/strong\u003e\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003eThe authors disclose no conflicts.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eF\u003c/strong\u003e\u003cstrong\u003einancial support statement\u003c/strong\u003e\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003eThis work was supported by the National Key Research and Development Program of China (No.2022YFC2303600), the National key clinical specialist construction Programs (2022001) and the National Natural Science Foundation of China under Grants No. 82371960.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank all individuals and patients who participated in this study. We also thank all pathologists who participated in the study. We thank AiMi Academic Services (www.aimieditor.com) for English language editing and review services.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe BJ-HepaGS code is available at https://github.com/zznn-ying/BJ-HepaGS.git\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBurki T. WHO\u0026apos;s 2024 global hepatitis report. Lancet Infect Dis. 2024; 24: e362-e363. doi: 10.1016/S1473-3099(24)00307-4. \u003c/li\u003e\n\u003cli\u003eJeng WJ, Papatheodoridis GV, Lok ASF. Hepatitis B. Lancet. 2023; 401: 1039-1052. doi: 10.1016/S0140-6736(22)01468-4. \u003c/li\u003e\n\u003cli\u003eWorld Health Organization. Global progress report on HIV, viral hepatitis and sexually transmitted infections, 2021. Accountability for the global health sector strategies 2016-2021: actions for impact[EB/OL]. 2021. https://www.who.int/publications/i/item/9789240027077.\u003c/li\u003e\n\u003cli\u003eHsu, YC., Huang, D.Q. \u0026amp; Nguyen, M.H. Global burden of hepatitis B virus: current status, missed opportunities and a call for action. Nat Rev Gastroenterol Hepatol. 2023; 20: 524-537. doi: 10.1038/s41575-023-00760-9. \u003c/li\u003e\n\u003cli\u003eMcMahon BJ. Meeting the WHO and US Goals to Eliminate Hepatitis B Infection by 2030: Opportunities and Challenges. Clin Liver Dis (Hoboken). 2018; 12: 29-32. doi: 10.1002/cld.733. \u003c/li\u003e\n\u003cli\u003eWorld Health Organization. Global health sector strategy on viral hepatitis 2016- 2021: Towards ending viral hepatitis[EB/OL]. https://apps.who.int/iris/bitstream/10665/246177/1/WHO-HIV-2016.06-eng,pdf?ua= 1. https://apps.who.int/iris/bitstream/10665/246177/1/WHO-HIV-2016.06-eng\u003c/li\u003e\n\u003cli\u003eZhang X, Jie Y, Wan Z, et al. Prognostic Value of Inflammatory Indicators in Chronic Hepatitis B Patients With Significant Liver Fibrosis: A Multicenter Study in China. Front Pharmacol. 2021; 12: 653751. doi: 10.3389/fphar.2021.653751.\u003c/li\u003e\n\u003cli\u003eBera C, Hamdan-Perez N, Patel K. Non-Invasive Assessment of Liver Fibrosis in Hepatitis B Patients. J Clin Med. 2024; 13 :1046. doi: 10.3390/jcm13041046. \u003c/li\u003e\n\u003cli\u003eTerrault NA, Lok ASF, McMahon BJ, et al. Update on prevention, diagnosis, and treatment of chronic hepatitis B: AASLD 2018 hepatitis B guidance. Hepatology. 2018; 67: 1560-1599. doi: 10.1002/hep.29800.\u003c/li\u003e\n\u003cli\u003eEuropean Association for the Study of the Liver. EASL 2017 Clinical Practice Guidelines on the management of hepatitis B virus infection. J Hepatol. 2017; 67: 370-398. doi: 10.1016/j.jhep.2017.03.021.\u003c/li\u003e\n\u003cli\u003eWorld Health Organization. Guidelines for the prevention,diagnosis, care and treatment for people with chronic hepatitis B infection[EB/OL]. (2024-03-29)[2024-12-18]. https://www.who.int/publications/i/item/9789240090903.\u003c/li\u003e\n\u003cli\u003eMani H, Kleiner DE. Liver biopsy findings in chronic hepatitis B. Hepatology. 2009; 49: S61-71. doi: 10.1002/hep.22930. \u003c/li\u003e\n\u003cli\u003eTong XF, Wang QY, Zhao XY, et al. Histological assessment based on liver biopsy: the value and challenges in NASH drug development. Acta Pharmacol Sin. 2022; 43: 1200-1209. doi: 10.1038/s41401-022-00874-x.\u003c/li\u003e\n\u003cli\u003eCastera L, Hartmann DJ, Chapel F, et al. Serum laminin and type IV collagen are accurate markers of histologically severe alcoholic hepatitis in patients with cirrhosis. J Hepatol. 2000; 32: 412-418. doi: 10.1016/s0168-8278(00)80391-8.\u003c/li\u003e\n\u003cli\u003eKishanifarahani Z, Ahadi M, Kazeminejad B, et al. Inter-observer Variability in Histomorphological Evaluation of Non-neoplastic Liver Biopsy Tissue and Impact of Clinical Information on Final Diagnosis in Shahid Beheshti University of Medical Sciences Affiliated Hospitals. Iran J Pathol. 2019; 14: 243-247. doi: 10.30699/ijp.2019.99566.1985.\u003c/li\u003e\n\u003cli\u003eNeuberger J, Cain O. The Need for Alternatives to Liver Biopsies: Non-Invasive Analytics and Diagnostics. Hepat Med. 2021; 13: 59-69. doi: 10.2147/HMER.\u003c/li\u003e\n\u003cli\u003eMalik S, Das R, Thongtan T, et al. AI in Hepatology: Revolutionizing the Diagnosis and Management of Liver Disease. J Clin Med. 2024; 13: 7833. doi: 10.3390/jcm13247833. \u003c/li\u003e\n\u003cli\u003eKishanifarahani Z, Ahadi M, Kazeminejad B, et al. Inter-observer Variability in Histomorphological Evaluation of Non-neoplastic Liver Biopsy Tissue and Impact of Clinical Information on Final Diagnosis in Shahid Beheshti University of Medical Sciences Affiliated Hospitals. Iran J Pathol. 2019; 14: 243-247. doi: 10.30699/ijp.2019.99566.1985. Epub 2019 Aug 1.\u003c/li\u003e\n\u003cli\u003eSeven. İ., Bayram, D., Arslan. H, et al. Predicting hepatocellular carcinoma survival with artificial intelligence. Sci Rep.2025; 15, 6226. doi: 10.1038/s41598-025-90884-6. \u003c/li\u003e\n\u003cli\u003ePerincheri S, Levi AW, Celli R. et al. An independent assessment of an artificial intelligence system for prostate cancer detection shows strong diagnostic accuracy. Mod. Pathol. 2021; 34, 1588-1595. doi: 10.1038/s41379-021-00794-x.\u003c/li\u003e\n\u003cli\u003eCalderaro, J., Ghaffari Laleh, N., Zeng. Q, et al. Deep learning-based phenotyping reclassifies combined hepatocellular-cholangiocarcinoma. Nat Commun. 2023; 14: 8290. doi: 10.1038/s41467-023-43749-3.\u003c/li\u003e\n\u003cli\u003eKalapala R, Rughwani H, Reddy DN. Artificial Intelligence in Hepatology- Ready for the Primetime. J Clin Exp Hepatol. 2023; 13: 149-161. doi: 10.1016/j.jceh.2022.06.009.\u003c/li\u003e\n\u003cli\u003eLu MY, Williamson DFK, Chen TY, et al. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. 2021; 5: 555-570. doi: 10.1038/s41551-020-00682-w.\u003c/li\u003e\n\u003cli\u003eChen RJ, Ding T, Lu MY, et al. Towards a general-purpose foundation model for computational pathology. Nat Med. 2024; 30: 850-862. doi: 10.1038/s41591-024-02857-3.\u003c/li\u003e\n\u003cli\u003eBedossa P, Darg\u0026egrave;re D, Paradis V. Sampling variability of liver fibrosis in chronic hepatitis C. Hepatology. 2003; 38: 1449-1457. doi: 10.1016/j.hep.2003.09.022.\u003c/li\u003e\n\u003cli\u003eRousselet MC, Michalak S, Dupre F, et al. Sources of variability in histological scoring of chronic viral hepatitis. Hepatology. 2005; 41: 257-264. doi: 10.1002/hep.20535.\u003c/li\u003e\n\u003cli\u003ePatil A, Salvatori R, Smith L, et al. Artificial intelligence-based reticulin proportionate area\u0026thinsp;-\u0026thinsp;a novel histological outcome predictor in hepatocellular carcinoma. Histopathology. 2023; 83: 512-525. doi: 10.1111/his.15001. \u003c/li\u003e\n\u003cli\u003eCalderaro J, Ghaffari Laleh N, Zeng Q, et al. Deep learning-based phenotyping reclassifies combined hepatocellular-cholangiocarcinoma. Nat Commun. 2023; 14: 8290. doi: 10.1038/s41467-023-43749-3.\u003c/li\u003e\n\u003cli\u003ePulaski H, Harrison SA, Mehta SS, et al. Clinical validation of an AI-based pathology tool for scoring of metabolic dysfunction-associated steatohepatitis. Nat Med. 2025; 31: 315-322. doi: 10.1038/s41591-024-03301-2. \u003c/li\u003e\n\u003cli\u003eZeng Q, Klein C, Caruso S, et al. Artificial intelligence-based pathology as a biomarker of sensitivity to atezolizumab-bevacizumab in patients with hepatocellular carcinoma: a multicentre retrospective study. Lancet Oncol. 2023; 24: 1411-1422. doi: 10.1016/S1470-2045(23)00468-0.\u003c/li\u003e\n\u003cli\u003eSanyal AJ, Loomba R, Anstee QM, et al. Utility of pathologist panels for achieving consensus in NASH histologic scoring in clinical trials: Data from a phase 3 study. Hepatol Commun. 2023; 8: e0325. doi: 10.1097/HC9.0000000000000325.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1. Clinical and histopathological characteristics of the training set and the validation sets.\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" align=\"left\" width=\"916\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePatient \u0026nbsp;characteristics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Set\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(n = 673)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation Set 1\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e(n = 300)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003ep\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e* value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePaired set (pre-treatment)\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(n = 50)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePaired set \u0026nbsp;(post-treatment)\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(n = 50)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003ep\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e\u003csup\u003e#\u003c/sup\u003e\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\" style=\"width: 916px;\"\u003e\n \u003cp\u003eDemographic parameters\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eMale sex, n(%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e509 (76%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e231 (77%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.644\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e27 (54%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e32 (64%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e---\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eAge, years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e44 (36,51)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e43.6 (36.0,51.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.747\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e39 (33.8, 48.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e38.0 (32.8, 47.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e---\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\" style=\"width: 916px;\"\u003e\n \u003cp\u003eSerolocical parameters\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eALT level,\u0026nbsp;U/L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e30.7 (21.0, 50.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e30.9 (21.0, 46.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.926\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e45.7 (33.0, 84.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e26.0 (18.9, 40.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eAST level,\u0026nbsp;U/L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e27.1 (22.0, 38.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e27.4 (23.0, 36.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.768\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e35.5 (26.9, 65.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e22.5 (20.0, 28.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eALB level,\u0026nbsp;U/L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e46.0 (43.0, 48.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e46.2 (43.1, 48.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.756\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e44.6 (40.8, 47.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e46.5 (45.0, 48.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eTBIL level,\u0026nbsp;\u0026mu;mol/L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e14.6 (10.7, 19.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e14.6 (10.9, 19.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.706\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e14.7 (12.0, 19.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e15.1 (10.5, 17.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e0.486\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eHBeAg (+), \u0026nbsp;%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e31%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e28%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.741\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e56%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e58%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e0.705\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eHBV DNA (+), %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e37%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e24%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003ePLT ,\u0026times;10\u003csup\u003e9\u003c/sup\u003e/L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e168.0 (129.0, 213.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e165.0 (124.0, 208.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.247\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e182.0 (140, 221.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e190.5 (150.5,224.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e0.011\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eLSM, kPa\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e8.9 (6.4, 12.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e8.3 (6.4, 10.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.078\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e9.9 (7.7, 15.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e6.6 (5.9, 8.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"3\" style=\"width: 447px;\"\u003e\n \u003cp\u003eNecroinflammation grade\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.814\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 326px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eG0, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e182 (27%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e86 (29%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eG0, 3 (6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e12 (24%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eG1, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e281 (42%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e114 (38%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eG1, 11 (22%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e23 (46%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eG2, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e154 (23%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e68 (23%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eG2, 18 (36%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e14 (28%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eG3/4, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e56 (8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e32 (10%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eG3/4, 18 (36%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e1 (2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"3\" style=\"width: 447px;\"\u003e\n \u003cp\u003eFibrosis stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 326px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eF0/1, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e279 (41%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e91 (30%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eF0/1, 19 (38%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e33 (66%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eF2, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e77 (11%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e36 (12%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eF2, 5 (10%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e2 (4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eF3, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e218 (32%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e120 (40%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eF3, 15 (30%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e12 (24%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eF4, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 154px;\"\u003e\n \u003cp\u003e99 (15%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003e53 (18%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eF4, 11 (22%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 166px;\"\u003e\n \u003cp\u003e3 (6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eNote: Data are presented as counts (percentages) or medians (with interquartile ranges, IQR) for continuous variables with skewed distributions. The p\u003csup\u003e*\u003c/sup\u003e value indicates differences between the training set and Validation Set 1, analyzed using the Mann-Whitney U test. The p\u003csup\u003e#\u0026nbsp;\u003c/sup\u003evalue indicates differences between the paired sets before and after treatment, analyzed using the Wilcoxon Signed-Rank Test.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Performance of the BJ-HepaGS model in histopathological grading and staging in independent validation set 1 and paired sets.\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" align=\"\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUROC (95%CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" style=\"width: 538px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation\u0026nbsp;set\u0026nbsp;1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003eInflammation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;G1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.890\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.957\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.720\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.91\u0026nbsp;(0.87-0.95)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;G2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.876\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.800\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.915\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.95\u0026nbsp;(0.92-0.97)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;G3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.946\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.718\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.973\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.98\u0026nbsp;(0.97-0.99)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003eFibrosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;F2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.850\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.895\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.748\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.89\u0026nbsp;(0.84-0.93)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;F3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.810\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.931\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.646\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.87\u0026nbsp;(0.83-0.91)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003eF4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.847\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.340\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.956\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.82\u0026nbsp;(0.76-0.88)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" style=\"width: 538px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation\u0026nbsp;paired\u0026nbsp;set\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003eInflammation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;G1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.910\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.953\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.667\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.92\u0026nbsp;(0.85-0.97)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;G2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.900\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.902\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.898\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.98\u0026nbsp;(0.96-0.99)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;G3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.940\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.842\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.963\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.96\u0026nbsp;(0.92-0.99)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003eFibrosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;F2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.780\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.875\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.692\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.90\u0026nbsp;(0.84-0.95)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003e\u0026ge;F3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.810\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.927\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.729\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.92\u0026nbsp;(0.87-0.97)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 119px;\"\u003e\n \u003cp\u003eF4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 98px;\"\u003e\n \u003cp\u003e0.870\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e0.571\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 92px;\"\u003e\n \u003cp\u003e0.919\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.91\u0026nbsp;(0.82-0.98)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Concordance of the BJ-HepaGS model and expert consensus\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"585\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHistology \u0026nbsp; \u0026nbsp; \u0026nbsp;features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eICC of \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; AI model \u0026nbsp;vs. Consensus\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eICC of\u0026nbsp;\u003cbr\u003e\u0026nbsp;Senior experts scores vs. Consensus\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eICC of \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Junior experts scores vs. Consensus\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 279px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation\u0026nbsp;set\u0026nbsp;1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInflammation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.824\u0026nbsp;(0.784-0.857)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.961\u0026nbsp;(0.952-0.969)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.763\u0026nbsp;(0.711-0.807)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;G1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.721\u0026nbsp;(0.662-0.771)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.984\u0026nbsp;(0.980-0.987)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.477\u0026nbsp;(0.385-0.560)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;G2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.721\u0026nbsp;(0.661-0.771)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.917\u0026nbsp;(0.897-0.933)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.755\u0026nbsp;(0.702-0.800)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;G3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.712\u0026nbsp;(0.652-0.764)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.930\u0026nbsp;(0.913-0.944)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.756\u0026nbsp;(0.703-0.801)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFibrosis\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.681\u0026nbsp;(0.615-0.737)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.931\u0026nbsp;(0.915-0.945)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.912\u0026nbsp;(0.891-0.930)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;F2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.644\u0026nbsp;(0.573-0.706)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.914\u0026nbsp;(0.893-0.930)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.905\u0026nbsp;(0.883-0.924)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;F3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.613\u0026nbsp;(0.537-0.679)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.888\u0026nbsp;(0.861-0.909)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.863\u0026nbsp;(0.831-0.889)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003eF4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.369\u0026nbsp;(0.267-0.463)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.829\u0026nbsp;(0.790-0.862)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.804\u0026nbsp;(0.760-0.840)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 279px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation\u0026nbsp;paired\u0026nbsp;set\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInflammation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.865\u0026nbsp;(0.805-0.907)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.942\u0026nbsp;(0.915-0.960)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.831\u0026nbsp;(0.758-0.883)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;G1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.637\u0026nbsp;(0.505-0.741)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.962\u0026nbsp;(0.944-0.974)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.618\u0026nbsp;(0.480-0.726)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;G2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.800\u0026nbsp;(0.716-0.861)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.841\u0026nbsp;(0.772-0.890)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.574\u0026nbsp;(0.427-0.692)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;G3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.805\u0026nbsp;(0.723-0.865)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.934\u0026nbsp;(0.903-0.955)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.754\u0026nbsp;(0.655-0.827)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFibrosis\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.682\u0026nbsp;(0.562-0.775)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.930\u0026nbsp;(0.897-0.952)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.910\u0026nbsp;(0.870-0.939)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;F2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.574\u0026nbsp;(0.427-0.692)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.920\u0026nbsp;(0.883-0.945)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.840\u0026nbsp;(0.771-0.889)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003e\u0026ge;F3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.647\u0026nbsp;(0.517-0.748)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.938\u0026nbsp;(0.909-0.958)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.896\u0026nbsp;(0.850-0.929)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 108px;\"\u003e\n \u003cp\u003eF4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.476\u0026nbsp;(0.309-0.614)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e0.812\u0026nbsp;(0.733-0.870)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 172px;\"\u003e\n \u003cp\u003e0.722\u0026nbsp;(0.614-0.804)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. The BJ-HepaGS model improved the consistency of scores between senior and junior experts.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"599\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHistology \u0026nbsp; \u0026nbsp; \u0026nbsp;features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eICC of\u0026nbsp;\u003cbr\u003e\u0026nbsp;Senior vs.\u0026nbsp;\u003cbr\u003e\u0026nbsp;Junior experts scores\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eICC of\u0026nbsp;\u003cbr\u003e\u0026nbsp;Senior vs. Junior experts scores refined by AI model\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"3\" style=\"width: 599px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation\u0026nbsp;set\u0026nbsp;1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInflammation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.730\u0026nbsp;(0.673-0.779)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.818\u0026nbsp;(0.777-0.852)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u0026ge;G1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e0.476\u0026nbsp;(0.383-0.559)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e0.460\u0026nbsp;(0.366-0.545)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u0026ge;G2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e0.684\u0026nbsp;(0.619-0.740)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e0.835\u0026nbsp;(0.798-0.867)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u0026ge;G3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e0.659\u0026nbsp;(0.590-0.719)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e0.822\u0026nbsp;(0.782-0.856)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFibrosis\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.852\u0026nbsp;(0.817-0.880)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.904\u0026nbsp;(0.881-0.922)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u0026ge;F2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e0.819\u0026nbsp;(0.778-0.853)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e0.860\u0026nbsp;(0.827-0.886)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e\u0026ge;F3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e0.766\u0026nbsp;(0.715-0.809)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e0.889\u0026nbsp;(0.863-0.911)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003eF4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 241px;\"\u003e\n \u003cp\u003e0.718\u0026nbsp;(0.659-0.769)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 254px;\"\u003e\n \u003cp\u003e0.793\u0026nbsp;(0.747-0.831)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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":"Chronic viral hepatitis, Necroinflammation, Fibrosis, Cirrhosis, METAVIR system, Hematoxylin and Eosin (H\u0026E) staining, Whole slide images","lastPublishedDoi":"10.21203/rs.3.rs-7440962/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7440962/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground and aims:\u003c/h2\u003e\u003cp\u003eChronic hepatitis B (CHB) is a global challenge, with histological assessment affected by observer variability. We aim to develop and validate BJ-HepaGS, an AI model for consistent evaluation.\u003c/p\u003e\u003ch2\u003eMethods:\u003c/h2\u003e\u003cp\u003eBJ-HepaGS was developed using hematoxylin and eosin (H\u0026amp;E)-stained whole-slide images (WSIs) from CHB patients across multiple hospitals. Model performance was validated using area under the curve (AUC) and intraclass correlation coefficient (ICC) metrics.\u003c/p\u003e\u003ch2\u003eResults:\u003c/h2\u003e\u003cp\u003eThe BJ-HepaGS trained on 673 WSIs (627,703 patches) and validated on independent (300 H\u0026amp;E-WSIs, 331,037 patches) and paired cohorts (n\u0026thinsp;=\u0026thinsp;100 H\u0026amp;E-WSIs, 52,271 patches). The independent set achieved area under the curve (AUC) values of 0.91\u0026ndash;0.98 for grading and 0.85\u0026ndash;0.91 for staging, with strong consistency versus expert consensus (ICC\u0026thinsp;=\u0026thinsp;0.824 and 0.681, respectively). BJ-HepaGS distinguished fibrosis stage F0-1 vs. F2-4 (87.6% accuracy) and cirrhosis (F0-3 vs. F4; 86.0% accuracy), and reliably assessed inflammation improvement (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.885) and fibrosis regression (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.388) in pre- and post-treatment paired samples. With the assistance of AI, the consistency between senior and junior expert interpretations on inflammation and fibrosis were significantly enhanced (both \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e\u003ch2\u003eConclusions:\u003c/h2\u003e\u003cp\u003eBJ-HepaGS addresses a key gap in CHB care by providing reproducible, objective histopathological interpretation, supporting standardized diagnosis and improved clinical management.\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence Model for Accurate Grading and Staging of Chronic Hepatitis B: Development and Validation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-17 19:14:56","doi":"10.21203/rs.3.rs-7440962/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":"e21ae1fa-8b07-43f5-9da5-9ba5b61d94f3","owner":[],"postedDate":"September 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":54112225,"name":"Health sciences/Medical research/Experimental models of disease"},{"id":54112226,"name":"Health sciences/Gastroenterology/Gastrointestinal models"}],"tags":[],"updatedAt":"2025-10-14T15:10:43+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-17 19:14:56","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7440962","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7440962","identity":"rs-7440962","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}