A nomogram model integrating ultrasound-based multimodal radiomics features and clinical indexes for diagnosing significant hepatic fibrosis in AILD patients

doi:10.21203/rs.3.rs-7801078/v1

Download PDF

Research Article

A nomogram model integrating ultrasound-based multimodal radiomics features and clinical indexes for diagnosing significant hepatic fibrosis in AILD patients

https://doi.org/10.21203/rs.3.rs-7801078/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 19 Dec, 2025

Read the published version in Abdominal Radiology →

You are reading this latest preprint version

Objective

To develop a prediction model combining radiomics features from 2D ultrasound (2D-US) and shear wave elastography (SWE) with clinical indicators for assessing significant hepatic fibrosis (S2–4) in autoimmune liver diseases (AILDs).

Methods

A total of 147 biopsy-confirmed AILD patients were classified into non-significant (S0–1, n = 44) and significant fibrosis (S2–4, n = 103) groups based on Scheuer’s classification, and randomly divided into training (n = 102) and validation (n = 45) cohorts. Radiomics features with interclass correlation coefficient > 0.75 were selected. Ten non-zero coefficient features were identified using least absolute shrinkage and selection operator (LASSO) regression. Six machine learning algorithms were evaluated. A nomogram integrating optimal radiomics features and clinical indexes was developed and assessed via ROC, calibration curve, and decision curve analysis.

Results

Logistic regression showed the best performance. Platelet count (PLT, OR = 0.991) and shear wave velocity (Vs, OR = 3.563) were independent predictors (P < 0.05). The combined nomogram achieved AUCs of 0.860 (training) and 0.912 (validation), significantly outperforming radiomics-only models, FIB-4, and APRI (P < 0.05). Calibration and decision curves indicated high clinical utility.

Conclusion

The nomogram integrating 2D-US/SWE radiomics and clinical indexes facilitates non-invasive diagnosis of significant fibrosis in AILDs, thus providing a more reliable quantitative tool for individualized assessments and clinical decision-making.

Advances in knowledge

This study develops the first nomogram combining multimodal ultrasound radiomics and clinical indexes for noninvasive diagnosis of significant hepatic fibrosis in autoimmune liver diseases, demonstrating superior diagnostic performance.

Radiomics

autoimmune liver disease

machine learning

2D ultrasound

shear wave elastography

Autoimmune liver diseases (AILDs) include autoimmune hepatitis (AIH), primary biliary cholangitis (PBC), primary sclerosing cholangitis (PSC), and overlap syndrome (OS) with characteristics of two or more types of the above-mentioned diseases Recent epidemiological studies indicate a rising global incidence of AILDs in recent years [1]. AILDs usually begins insidiously with non-specific clinical symptoms [2]. The pathological mechanism involves abnormal infiltration of immune cells and persistent bile duct inflammation, which may eventually progress to hepatic fibrosis, hepatic cirrhosis, and even hepatic failure or hepatocellular carcinoma, thereby seriously affecting the prognosis of patients [3, 4]. Therefore, early identification of the degree of fibrosis in AILDs is of significant importance for treatment and prognosis improvement of this disease.

At present, liver biopsy remains the gold standard for evaluation of the degree of hepatic fibrosis. However, this technique is an invasive procedure with operational risks (bleeding, infection) and sampling errors, making it difficult to dynamically monitor disease prognosis during clinical management [5]. Among non-invasive detection methods, the serum biomarkers such as aspartate aminotransferase to platelet ratio index(APRI) and the fibrosis index based on four factors(FIB-4) have certain values, but their diagnostic efficacies are limited [6, 7]. In imaging techniques, the ultrasound-based elastography such as 2D-SWE can quantitatively assess the liver stiffness measurement (LSM), which is of great value in monitoring therapeutic efficacy and evaluating prognosis [8–10]. However, the detection results are easily affected by factors such as operator dependence, patient obesity, aminotransferase level, and hepatic steatosis, which may lower the diagnostic consistency [11, 12].

With the development of artificial intelligence technology, the collaborative application of radiomics and machine learning provides a new research direction for the non-invasive evaluation of hepatic fibrosis and obtains better diagnostic performance [13–15]. In this study, the radiomics features extracted from 2D-ultrasound (2D-US) and shear-wave elastography (SWE) images were integrated with the clinical indexes to construct a novel prediction model and explore its value in the diagnosis of significant hepatic fibrosis in AILDs, thus providing new ideas for early clinical diagnosis and treatment of AILDs.

Ethical statement

This study complies with the Declaration of Helsinki 1975, and has received approval from the Ethics Committee in our hospital (EK2024058). All patients had signed a written informed consent form before undergoing hepatic biopsy.

Study subjects

In this study, a total of 147 patients with AILDs confirmed by liver biopsy, who were treated in our hospital from 2021 to 2024, were retrospectively analyzed (Fig.1). Inclusion criteria: (1) patients with an age of more than 18 years; (2) patients undergoing laboratory examination, 2D-US, and SWE 1 week before liver biopsy; (3) patients with definite inflammation grade (G) and fibrosis stage (S) confirmed by the histopathological examination. Exclusion criteria: (1) patients with unreliable pathological findings; (2) patients with missed clinical or ultrasound data;(3) patients with concurrent chronic liver diseases caused by other causes (e.g., viral hepatitis, drug-induced liver injury, alcoholic hepatic disease, and metabolic-associated fatty liver disease).

Instruments and methods for ultrasound examination

The diagnostic ultrasound system Aloka ARIETTA 850 (Hitachi Medical Systems, Japan) with an abdominal probe was used at frequencies of 1-6 MHz to perform ultrasound examinations. Before the examination, the patient was required to fast for 4 hours, then the patient lay in a supine position, with both hands raised above the head. and the upper abdomen was adequately exposed for 2D-US image acquisition; SWE images were collected when the patients were holding their breath during calm respiration. SWE indexes included shear wave velocity (Vs), liver fibrosis index (LFI), inflammation activity-related index (A index), and attenuation in the liver (ATT). All liver imaging examinations (including SWE) were performed by senior physicians with over 10 years of experience in abdominal imaging.

Clinical data of patients

The clinical data of patients included age, gender, alanine aminotransferase (ALT), aspartate aminotransferase (AST), globulin (GLOB), total bilirubin (Bil), albumin (ALB), and platelet (PLT).

Pathological examination

The patients underwent ultrasound-guided liver biopsy on the same day as the ultrasound examination. An automated biopsy gun (16G, GMT Medical) was used to collect samples from the right lobe of the liver. The hepatic fibrosis was staged using a modified Scheuer scoring system: S0 indicated no fibrosis; S1 indicated fibrous portal expansion; S2 indicated formation of occasional fibrous septage with preserved hepatic lobular structures; S3 indicated many fibrous septage with disrupted hepatic lobular structures and no cirrhosis; S4 indicated early hepatic cirrhosis or definite hepatic cirrhosis. Of which, S0-S1 was defined as no significant hepatic fibrosis, while S2-S4 was defined as significant hepatic fibrosis.

Radiomics workflow

Image segmentation and feature extraction

The original images were imported into ITK-SNAP 4.0（http://www.itksnap.org), and two researchers used ITK-SNAP 4.0 to independently delineate two regions-of-interest (ROIs) in 2D-US images (entire left lateral external lobe of liver) and SWE images (entire elastography sampling box) respectively for each patient (Fig. 2). The radiomics features were extracted from two ROIs in each patient, including first-order statistics, textural features, and wavelet-transformed features [16]. In-class correlation coefficients (ICCs) were used to evaluate the inter-observer and intra-observer consistency, and the features with an ICC more than 0.75 were retained for subsequent analysis. The most suitable features with non-zero coefficients were screened using the least absolute shrinkage and selection operator (LASSO) regression,

Model construction and validation

Based on the data obtained during the feature selection phase, the scikit-learn toolkit, which included six machine learning algorithms such as Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Extra Trees (ET), extreme Gradient Boosting (Boost) and Light Gradient Boosting Machine (Light), were used to construct a machine learning model in this study[17-22].In this study, in order to identify an optimal machine learning model, the area under the curve(AUC) was used as the main evaluation index, and the accuracy, F1 score, sensitivity and specificity were used as additional evaluation indexes [23]. The univariate and multivariate logistic regression analyses of SWE features and clinical indexes in the training cohort were performed to screen the variables with p < 0.05, which may be determined as independent risk factors for hepatic fibrosis. Use Shapley Additive Explanations (SHAP) to explain ML models and analyze the features of the model. A nomogram model was constructed by integrating the optimal radiomics features and clinical indexes, and the performance of the constructed model was evaluated and validated in the validation cohort. Finally, the clinical application value of this model was evaluated by using the calibration curves and decision curve analysis (DCA).

Statistical analysis

Statistical analyses were performed using the software’s such as SPSS 29.0 (http://www.spss.com.hk/), GraphPad Prism 8.0 (https://www.graphpad.com/), and Python (https://www.python.org/). Continuous variables were analyzed using either the t-test or the Mann-Whitney U test depending on whether they followed a normal distribution, and they were summarized by means and standard deviations or medians. Multivariable analysis was conducted using a logistic regression (LR) model, and odds ratios (ORs) were reported. A p-value of < 0.05 indicated that the difference was statistically significant.

Baseline characteristics of patients

The clinical baseline characteristics of the patients are shown in Table 2. A total of 147 patients with AILDs were included in the study, including 21 males and 126 females, with an age range of 48-61 years, of whom, 15 patients had AIH, 45 patients had AIH-PBC, 2 patients had drug-induced autoimmune Hepatitis（DIAIH，83 patients had PBC and 2 patients had PSC; 44 patients had S0-1 hepatic fibrosis, and 103 patients had S2-4 hepatic fibrosis. Statistical analyses showed no significant differences in demographic data, clinical and imaging indexes between the two groups (P > 0.05).

Extraction and selection of radiomics features

We extracted 1561 radiomics features respectively from the ROIs of both 2D-US images and SWE images in each patient, and the quantified radiomics features were normalized using Z-score normalization. The radiomics features of 1302 2D-US images and 1078 SWE images showed good consistency (ICC > 0.75) based on the results of intra- and inter-observer consistency analysis performed by two researchers. Pearson’s correlation test and principal component analysis (PCA) of the radiomics features were conducted, and 10 non-zero coefficient radiomics features (involving 6 2D-US images and 4 SWE images) were selected using the LASSO regression and mean squared error (MSE) (Fig. 3).

Selection of optimal machine learning algorithm

In this study, six machine learning algorithms were used to construct a multimodal radiomics model integrating 2D-US and SWE features in the training cohort, which was validated in the validation cohort (Table 1). The LR model exhibit better predictive performance in the validation cohort, with an AUC of 0.838 (95% CI: 0.686- 0.991). Therefore, the LR model was used as the core classification algorithm in the single-modal models comparing S0-1 and S2-4 hepatic fibrosis. A total of three radiomics models were constructed, namely the 2D-US-based radiomics model (using only 2D-US features); the SWE based radiomics model (using only SWE features) and the 2D-US/SWE model (using both 2D-US and SWE features).

SHAP analysis was employed to visualize the predictive performance of the model integrating 2D-US and SWE features (Fig. 4 A). The SHAP-based feature importance analysis demonstrated that liver fibrosis staging prediction predominantly depended on the synergistic effects of multimodal radiomics features. Notably, morphological characteristics (original_shape_Maximum2DDiameterRow) and wavelet-transformed texture features (software’s, software’s) exhibited the strongest positive influence on the model’s output. Figure 4 B illustrates a representative clinical case of AILDs with significant fibrosis, where blue and red arrows denote features that either reduce (blue) or enhance (red) the probability of significant liver fibrosis.

Selection of clinical indexes

The univariate and multivariate logistic regression analyses were performed to explore the correlations between clinical indexes and the stage of hepatic fibrosis (Table 3). The univariate analysis revealed that age, ALB, GLOB, LFI, A index, Vs, PLT, ALT, AST and ATT were significantly correlated with the stage of hepatic fibrosis (all P < 0.05). The multivariate analysis revealed that PLT (OR=0.991, 95% CI: 0.985–0.996, p=0.008) and Vs (OR=3.563, 95% CI: 1.259–10.085, p=0.045) were independent predictors for S2-4 hepatic fibrosis. A nomogram model integrating clinical indexes (Vs and PLT) and multimodal radiomics (2D-US/SWE) features were constructed (Fig. 5 A).

Model construction and evaluation

Table 4 showed the diagnostic performances of seven models in differentiating between S0-1 and S2-4 hepatic fibrosis in two different cohorts. Among them, the clinical-radiomics nomogram model, which integrated multimodal radiomics features (2D-US/SWE features) and clinical indexes (Vs and PLT), demonstrated the best diagnostic efficacy (Fig. 5). The results of the validation cohort showed that the AUC of the clinical-radiomics nomogram model was 0.912 (95% CI: 0.823-1.000), which was significantly higher than those of the multimodal (2D-US/SWE) radiomics model (0.838), Vs (0.818), 2D-US radiomics model (0.78), SWE radiomics model (0.755), FIB-4 (0.77) and APRI (0.748). The clinical-radiomics nomogram model also exhibited superior comprehensive performance: the accuracy (82.2%), sensitivity (81.8%), specificity (83.3%), and F1 score (0.871) were all at the leading level, suggesting a greater advantage in balancing diagnostic sensitivity and specificity (Table 4). The calibration curve analysis showed that the clinical-radiomics nomogram model demonstrated a good consistency between the predicted and actual values in both groups. Decision curve analysis of validation cohort showed that the clinical-radiomics nomogram model exhibited a superior net benefit within the high-risk threshold range from 1:100 to 3:2, and its standardized net benefit was significantly higher than those of other models with in a threshold range from 0.2 to 1.0 (Fig. 6).

The results of this study indicate that PLT and Vs are independent predictive factors for significant hepatic fibrosis in patients with AILDs. Among the six machine learning algorithms used in the study, the LR algorithm demonstrated good performance in predicting significant hepatic fibrosis. PLT, Vs, and LR algorithm-based multimodal radiomics features were combined in this study to construct a non-invasive nomogram prediction model, which aimed at predicting significant hepatic fibrosis in patients with AILDs, and this nomogram model exhibited superior diagnostic efficacy in both the training and validation cohorts.

In this study, PLT was an independent predictor of significant hepatic fibrosis, this result is consistent with those of the above-mentioned studies, which may explain that there is a correlation between hypersplenism and thrombocytopenia in the process of hepatic fibrosis [24,25]. Vs can also be used as an independent predictor of significant hepatic fibrosis, and we can indirectly assess tissue hardness by measuring the speed at which the shear waves travel through tissue, thus helping to assess the stage of hepatic fibrosis [26]. In a study of 114 patients with AILDs, Zeng et al. used 2D-SWE to predict significant hepatic fibrosis, with an AUC of 0.85 [27]. In this study, Vs was used to assess the significant hepatic fibrosis, with an AUC of 0.712 in the training cohort and 0.818 in the validation cohort, respectively. Considering that different ultrasonic instruments adopt different elasticity measurement methods and pathological scoring criteria, there may be a certain degree of variability. In addition, the patients with AILDs often have active hepatic inflammation, elevated aminotransferase level, and cholestasis, which often lead to bias in LSM value [12].

Therefore, we introduced radiomics and machine learning to enhance the diagnostic efficacy of the model. As a computer-aided quantitative analysis method, the radiomics can extract high-dimensional image data from the ROI using machine learning algorithms and convert them into radiomics features with pathological relevance [28,29]. At present, the "radiomics + machine learning" analytical approach has become a mainstream solution for medical image analysis [22]. The study by Xu et al. demonstrates that "radiomics + machine learning" can learn from image data, thus reducing interference from subjective factors and ensuring the objectivity and reliability of prediction results [30]. A radiomics model constructed by Zhao et al. based on the SVM algorithm exhibited an excellent ability to discriminate between mild and severe hepatic fibrosis induced by Schistosoma japonicum infection [19]. In this study, six machine learning algorithms (LR, Boost, SVM, RF, ET, and Light) were used for modeling and analysis of the extracted radiomics features. Among them, the LR algorithm had the best overall performance in the validation cohort (AUC = 0.838). Therefore, LR was selected as the core algorithm for the single-modal models. The advantage of this algorithm is that it combines an L2 normalization constraint with maximum likelihood function and adopts stochastic gradient descent optimization, which can effectively balance between the prediction accuracy of binary classification problems and the model complexity [24].

Some studies have confirmed [15,19] that the radiomics has a higher value in the diagnosis and staging of diffuse liver diseases, but the performance of the method based on single-modal US images in hepatic fibrosis staging is limited. At present, the multimodal ultrasound radiomics model that combine gray-scale ultrasonography with elastography (sound touch elastography, STE) can significantly increase the sensitivity and specificity of hepatic fibrosis assessment. For example, Ge et al. utilized a combination of STE and 2D-US to enhance the diagnostic performance for renal tubular interstitial fibrosis in chronic kidney disease [31]. Therefore, we used a multimodal radiomics model integrating both 2D-US and SWE features to reduce the impact of inflammation on the model.

The AUC of this multimodal nomogram model was 0.860 in the training cohort and 0.912 in validation cohorts, respectively, which was significantly better than those of single-modal model, FIB-4 and APRI models (Delong’s test, P < 0.05). This result suggests that this multimodal radiomics model can enhance classification performance through complementary information. This is consistent with the findings of the studies by Xue et al. A multimodal approach shows better performance compared with a single-modal approach, indicating the multimodal approach can carry more diagnostic information[32].The superiority of this combined nomogram model stems from the integration of multidimensional data: on one hand, the quantitative information on liver stiffness provided by SWE was included; on the other hand, the independent predictors such as PLT and Vs selected in the multifactor analysis were integrated, thereby the informational limitation of single imaging or serological index was overcome, and the diagnostic efficacy for significant hepatic fibrosis was enhanced.

Furthermore, SHAP-based interpretation of the 2D_SWE model identified original_

shape_Maximum2DDiameterRow, wavelet_LHH_gldm_SmallDependenceHighGray

LevelEmphasis and wavelet_HHH_gldm_DependenceVariance as the top three features exerting the strongest positive effects on model predictions. Specifically: The elevated SHAP value of original_shape_Maximum2DDiameterRow implies a potential association between macroscopic morphological irregularity and advancing fibrosis. The prominent contributions of wavelet-based features (wavelet_LHH_gldm_Small

DependenceHighGrayLevelEmphasis, wavelet_HHH_gldm_DependenceVariance) highlight the diagnostic relevance of microstructural heterogeneity in stratifying fibrosis stages. This finding not only corroborates the biological rationale underlying radiomics-based fibrosis evaluation but also offers clinically interpretable, quantifiable insights into the model’s decision-making process.

There are some limitations in this study. First, the enrolled patients included patients with various AILDs. There was an unbalanced sample size among patients with different AILDs, which may affect the diagnostic accuracy of the combined prediction model constructed in this study, which should be validated in each AILD separately in the future studies. Second, this study only included 147 patients with AILDs, and the patients with S0-1 hepatic fibrosis accounted for a relatively small proportion. The sample size needs to be expanded in future studies to reduce overfitting during model construction. Third, this was a single-center study, and only one type of ultrasound diagnostic device was used. In the future, multi-center studies and different ultrasonic devices will be needed for experiments to construct more generalizable combined models. Fourth, this study only included the image data of 2D-US and SWE, and future studies can incorporate contrast-enhanced ultrasound images and splenic images for constructing more multimodal models. Finally, because AILDs had a prolonged course, the follow-up time for patients with AILDs was insufficient in this study, and the follow-up time should be extended in the future studies to explore the effect of the combined prediction model on the prognosis of patients with AILDs.

A novel nomogram prediction model has been constructed in this study by integrating multimodal radiomics features (2D-US and SWE) and clinical indexes (Vs and PLT), which can significantly enhance the performance of conventional ultrasound in diagnosing significant hepatic fibrosis in patients with AILDs, thus providing a more reliable quantitative tool for individualized assessments and clinical decision-making.

A inflammation activity-related index

AIH autoimmune hepatitis

AILDs autoimmune liver diseases

ALB albumin

ALT alanine aminotransferase

APRI aminotransferase to platelet ratio index

AST aspartate aminotransferase

ATT attenuation

ET Extra Trees

FIB-4 fibrosis index based on four factors

LFI liver fibrosis index

Light Light Gradient Boosting Machine

LR logistic regression

PBC primary biliary cholangitis

PLT platelet

PSC primary sclerosing cholangitis

RF Random Forest

SVM Support Vector Machine

SWE shear-wave elastography

Vs shear wave velocity

XGBoost extreme Gradient Boosting

Author Contribution

Zixian Wang: Writing – original draft, Validation, Investigation, Software, Data curation, Methodology, ConceptualizationQiying Yu: Writing – review & editing, Supervision, Data curationYanan Sun: Visualization, Formal analysis, Data curationYanlou Liang: Data curation, Software, InvestigationShanshan Chen: Methodology, Provision of study materials or patientsShuhui Xie: Resources, Provision of study materials or patientsJing Wu: Writing – review & editing, Supervision, Software, Project administration, Conceptualization

Chen H, Shen Y, Wu SD, Zhu Q, Weng CZ, Zhang J, et al. Diagnostic role of transient elastography in patients with autoimmune liver diseases: a systematic review and meta-analysis. World J Gastroenterol. 2023;29:5503–25.
Trivedi PJ, Hirschfield GM. Recent advances in clinical practice: epidemiology of autoimmune liver diseases. Gut. 2021;70:1989–2003.
Chinese Medical Association. Guidelines on the diagnosis and management of primary sclerosing cholangitis (2021). Chin J Hepatol. 2021;60:1050–74.
Chinese Society of Hepatology, Chinese Medical Association. Guidelines on the diagnosis and management of autoimmune hepatitis (2021). Chin J Intern Med. 2021;60:1038–49.
Sumida Y, Nakajima A, Itoh Y. Limitations of liver biopsy and non-invasive diagnostic tests for the diagnosis of nonalcoholic fatty liver disease/nonalcoholic steatohepatitis. World J Gastroenterol. 2014;20:475–85.
Wang Z, Zhou Y, Yu P, Liu Y, Mei M, Bian Z, et al. Retrospective evaluation of non-invasive assessment based on routine laboratory markers for assessing advanced liver fibrosis in chronic hepatitis B patients. Int J Gen Med. 2022;15:5159–71.
Liu Z, Li W, Zhu Z, Wen H, Li MD, Hou C, et al. A deep learning model with data integration of ultrasound contrast-enhanced micro-flow cines, B-mode images, and clinical parameters for diagnosing significant liver fibrosis in patients with chronic hepatitis B. Eur Radiol. 2023;33:5871–81.
Yan Y, Xing X, Lu Q, Wang X, Luo X, Yang L. Assessment of biopsy proven liver fibrosis by two-dimensional shear wave elastography in patients with primary biliary cholangitis. Digest Liver Dis. 2020;52:555–60.
Zhuang Y, Ding H, Zhang Y, Sun H, Xu C, Wang W. Two-dimensional shear-wave elastography performance in the noninvasive evaluation of liver fibrosis in patients with chronic hepatitis B: comparison with serum fibrosis indexes. Radiology. 2017;283:873–82.
Lee Y, Kim NH, Kang JH, Cho YS, Kim Y, An J, et al. Validation of diagnostic thresholds for compensated advanced chronic liver disease using supersonic shear imaging. Radiology. 2024;311:e232188.
Cassinotto C, Lapuyade B, Mouries A, Hiriart JB, Vergniol J, Gaye D, et al. Non-invasive assessment of liver fibrosis with impulse elastography: comparison of Supersonic Shear Imaging with ARFI and FibroScan. J Hepatol. 2014;61:550–7.
Li C, Dhyani M, Bhan AK, Grajo JR, Pratt DS, Gee MS, et al. Diagnostic performance of shear wave elastography in patients with autoimmune liver disease. J Ultras Med. 2019;38:103–11.
Zhou Z, Zhang Z, Gao A, Tai DI, Wu S, Tsui PH. Liver fibrosis assessment using radiomics of ultrasound homodyned-K imaging based on the artificial neural network estimator. Ultrason Imaging. 2022;44:229–41.
Yin Y, Yakar D, Dierckx RA, Mouridsen KB, Kwee TC, De Haas RJ. Combining hepatic and splenic CT radiomic features improves radiomic analysis performance for liver fibrosis staging. Diagnostics (Basel). 2022;12:550.
Park HC, Joo Y, Lee OJ, Lee K, Song TK, Choi C, et al. Automated classification of liver fibrosis stages using ultrasound imaging. BMC Med Imaging. 2024;24:36.
Van Griethuysen JJ, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77:e104–7.
Li R, Li L, Xu Y, Yang J. Machine learning meets omics: applications and perspectives. Brief Bioinform. 2022;23:bbab460.
Xia F, Wei W, Wang J, Wang Y, Wang K, Zhang C, et al. Ultrasound radiomics-based logistic regression model for fibrotic NASH. BMC Gastroenterol. 2025;25:66.
Guo Z, Zhao M, Liu Z, Zheng J, Gong Y, Huang L, et al. Feasibility of ultrasound radiomics based models for classification of liver fibrosis due to Schistosoma japonicum infection. PLoS Negl Trop Dis. 2024;18:e0012235.
Chen Z, Xu H, Li Z, Zhang Y, Zhou T, You W, et al. Random survival forest: applying machine learning algorithm in survival analysis of biomedical data. Chin J Prev Med. 2021;55:104–9.
Ma B, Meng F, Yan G, Yan H, Chai B, Song F. Diagnostic classification of cancers using extreme gradient boosting algorithm and multi-omics data. Comput Biol Med. 2020;121:103761.
Wen DY, Chen JM, Tang ZP, Pang JS, Qin Q, Zhang L, et al. Noninvasive prediction of lymph node metastasis in pancreatic cancer using an ultrasound-based clinicoradiomics machine learning model. Biomed Eng Online. 2024;23:56.
Hicks SA, Strümke I, Thambawita V, Hammou M, Riegler MA, Halvorsen PAL, et al. On evaluation metrics for medical applications of artificial intelligence. Sci Rep. 2022;12:5979.
Su M, He M, Liu JL, Yang WC, Wang JJ, Guo C, et al. The therapeutic effect of leucogen in treating alcoholic liver cirrhosis with thrombocytopenia or leukopenia. Ann Saudi Med. 2025;45:9–17.
Manea M, Mărunțelu I, Constantinescu I. A new assessment of two transferase-based liver enzymes in low-and high-fibrosis patients chronically infected with hepatitis B virus: a meta-analysis and pilot study. J Clin Med. 2024;13:3903.
Yokoo T, Kanefuji T, Suda T, Nagayama I, Hoshi T, Abe S, et al. Rational arrangement of measuring shear wave speed in the liver. World J Gastroenterol. 2019;25:2503–13.
Zeng J, Huang ZP, Zheng J, Wu T, Zheng RQ. Non-invasive assessment of liver fibrosis using two-dimensional shear wave elastography in patients with autoimmune liver diseases. World J Gastroenterol. 2017;23:4839–46.
Zwanenburg A, Vallières M, Abdalah MA, Aerts HJ, Andrearczyk V, Apte A, et al. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology. 2020;295:328–38.
Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278:563–77.
Xu H, Deng L, Tian R, Ma X. novel methods for oncologic imaging analysis: radiomics, machine learning, and artificial intelligence. Front Oncol. 2021;11:628310.
Ge XY, Lan ZK, Lan QQ, Lin HS, Wang GD, Chen J. Diagnostic accuracy of ultrasound-based multimodal radiomics modeling for fibrosis detection in chronic kidney disease. Eur Radiol. 2023;33:2386–98.
Xue LY, Jiang ZY, Fu TT, Wang QM, Zhu YL, Dai M, et al. Transfer learning radiomics based on multimodal ultrasound imaging for staging liver fibrosis. Eur Radiol. 2020;30:2973–83.

Table 1
Development of 2D_SWE radiomic models using six machine learning algorithms.
Cohort	Model	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	F1
Training	LR	0.745	0.825	0.735–0.915	0.714	0.812	0.893	0.565	0.794
	SVM	0.912	0.969	0.942–0.996	0.871	1.000	1.000	0.780	0.931
	RF	0.843	0.935	0.889–0.982	0.800	0.937	0.966	0.682	0.875
	ET	0.775	0.848	0.762–0.935	0.757	0.812	0.898	0.605	0.822
	Boost	0.961	0.995	0.986–1.000	0.957	0.969	0.985	0.912	0.971
	Light	0.873	0.901	0.836–0.967	0.929	0.750	0.890	0.828	0.909
Validation	LR	0.822	0.838	0.686–0.991	0.848	0.750	0.903	0.643	0.875
	SVM	0.711	0.740	0.569–0.911	0.727	0.667	0.857	0.471	0.787
	RF	0.444	0.644	0.469–0.819	0.242	1.000	1.000	0.324	0.390
	ET	0.622	0.730	0.563–0.896	0.515	0.917	0.944	0.407	0.667
	Boost	0.733	0.770	0.619–0.921	0.727	0.750	0.889	0.500	0.800
	Light	0.756	0.745	0.560–0.930	0.788	0.667	0.867	0.533	0.825
LR Logistic Regression; SVM Support Vector Machine; RF Random Forest; ET Extra Trees; Boost extreme Gradient Boosting; Light Light Gradient Boosting Machine; AUC Area Under the Curve; CI confidence interval; PPV Positive Predictive Value; NPV Negative Predictive Value.

Table 2
Baseline clinical characteristics of the training and validation cohorts.
Variables	Total (n = 147)	Validation (n = 45)	Training (n = 102)	P value
Gender (n, %)				0.584
Female	126 (86)	37 (82)	89 (87)
Male	21 (14)	8 (18)	13 (13)
Age (year)	54 (48, 61)	54 (51, 61)	54 (47, 59.75)	0.367
Pathologic (n, %)				0.828
AIH	15 (10)	4 (9)	11 (11)
AIH-PBC	45 (31)	13 (29)	32 (31)
DIAIH	2 (1)	0 (0)	2 (2)
PBC	83 (56)	28 (62)	55 (54)
PSC	2 (1)	0 (0)	2 (2)
Vs (m/s)	1.82 (1.46, 2.29)	1.83 (1.38, 2.33)	1.81 (1.49, 2.27)	0.737
ATT (dB/s)	0.53 (0.49, 0.6)	0.53 (0.5, 0.62)	0.54 (0.47, 0.6)	0.412
LFI (kPa)	3.05 ± 1.12	2.81 ± 1.08	3.16 ± 1.12	0.079
A	1.3 (1.03, 1.6)	1.22 (1.02, 1.6)	1.34 (1.04, 1.59)	0.436
ALT (U/L)	71 (35, 133)	66 (35, 107)	74.5 (35.5, 154.5)	0.265
AST (U/L)	66 (41, 108.5)	59 (39, 88)	69 (41.25, 112.25)	0.284
ALB (g/L)	38.5 (35.3, 41.1)	39 (35.5, 40.8)	38 (35.08, 41.1)	0.667
GLOB (g/L)	31.9 (27.4, 36.8)	30.2 (26.9, 35.7)	32.4 (27.9, 36.95)	0.181
TBIL (µmol/L)	18.3 (13.2, 27.05)	19 (13, 28.5)	18.2 (13.85, 24.87)	0.895
PLT (×10⁹/L)	177.03 ± 70.82	176.64 ± 78.75	177.2 ± 67.44	0.968
Fibrosis grade (n, %)				0.584
S0-S1	44 (30)	12 (27)	32 (31)
S2	43 (29)	12 (27)	31 (30)
S3	32 (22)	13 (29)	19 (19)
S4	28 (19)	8 (18)	20 (20)
AIH autoimmune hepatitis; PBC primary biliary cholangitis; PSC primary sclerosing cholangitis; AIH-PBC AIH-PBC overlap syndrome; DIAIH Drug-Induced Autoimmune Hepatitis; Vs shear wave velocity; LFI liver fibrosis index; A index inflammation activity-related index, ATT attenuation coefficient; ALB albumin; TBIL total bilirubin; ALT alanine aminotransferase; AST Aspartate aminotransferase; PLT platelet count; GLOB Globulin. Values are presented as number (%), mean (SD) or median (IQR).

Table 3
Univariate and multivariate logistic regression analysis of significant liver fibrosis in AILD patients.
Characteristic	Univariate analysis			Multivariate analysis
Characteristic	OR	95% CI	p value	OR	95% CI	p value
PLT	1.003	1.001–1.005	0.013	0.991	0.985–0.996	0.008
ALT	1.004	1.002–1.007	0.006	0.999	0.995–1.004	0.762
AST	1.006	1.002–1.009	0.005	1.002	0.997–1.006	0.475
TBIL	1.006	0.998–1.014	0.231
Age	1.016	1.010–1.023	< 0.001	1.027	0.982–1.073	0.33
ALB	1.02	1.011–1.029	< 0.001	0.961	0.908–1.015	0.235
GLOB	1.026	1.015–1.038	< 0.001	1.022	0.957–1.091	0.591
LFI	1.286	1.153–1.435	< 0.001	0.904	0.615–1.330	0.668
A	1.484	1.160–1.900	0.009	0.655	0.302–1.420	0.369
VS	1.638	1.354–1.982	< 0.001	3.563	1.259–10.085	0.045
Gender	3.333	1.129–9.845	0.067
ATT	4.627	2.401–8.917	< 0.001	2.712	0.056–131.499	0.672
Vs shear wave velocity; LFI liver fibrosis index; A inflammation-related index, ATT attenuation coefficient; ALB albumin; TBIL total bilirubin; ALT alanine aminotransferase; AST Aspartate aminotransferase; PLT platelet count; GLOB Globulin; OR odds ratio; CI confidence interval.

Table 4
Comparison of predictive performance for significant fibrosis among different.
Cohort	Model	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	F1
Training	nomogram	0.833	0.860	0.780–0.940	0.871	0.750	0.884	0.727	0.878
	2D_SWE	0.745	0.825	0.735–0.915	0.714	0.812	0.893	0.565	0.794
	2D	0.676	0.771	0.675–0.867	0.600	0.844	0.894	0.491	0.718
	SWE	0.765	0.721	0.603–0.839	0.871	0.531	0.803	0.654	0.836
	FIB-4	0.529	0.691	0.587–0.795	0.329	1.000	1.000	0.392	0.485
	APRI	0.676	0.604	0.487–0.722	0.800	0.406	0.747	0.481	0.772
Validation	nomogram	0.822	0.912	0.823–1.000	0.818	0.833	0.931	0.625	0.871
	2D_SWE	0.822	0.838	0.686–0.991	0.848	0.750	0.903	0.643	0.875
	2D	0.756	0.780	0.628–0.933	0.788	0.667	0.867	0.533	0.825
	SWE	0.778	0.755	0.570–0.940	0.788	0.750	0.897	0.562	0.839
	FIB-4	0.711	0.770	0.622–0.919	0.667	0.833	0.917	0.476	0.772
	APRI	0.711	0.748	0.607–0.888	0.636	0.917	0.955	0.478	0.764
AUC Area Under the Curve; CI confidence interval; PPV Positive Predictive Value; NPV Negative Predictive Value.

No competing interests reported.

supplementarymaterials.docx

Download PDF

Journal Publication

published 19 Dec, 2025

Read the published version in Abdominal Radiology →

Editorial decision: Revision requested
27 Oct, 2025
Reviews received at journal
22 Oct, 2025
Reviews received at journal
20 Oct, 2025
Reviewers agreed at journal
09 Oct, 2025
Reviewers agreed at journal
09 Oct, 2025
Reviewers invited by journal
08 Oct, 2025
Editor assigned by journal
08 Oct, 2025
Submission checks completed at journal
08 Oct, 2025
First submitted to journal
07 Oct, 2025

You are reading this latest preprint version

A nomogram model integrating ultrasound-based multimodal radiomics features and clinical indexes for diagnosing significant hepatic fibrosis in AILD patients

Status:

Journal Publication

Version 1

Abstract

Objective

Methods

Results

Conclusion

Advances in knowledge

Figures

Introduction

Material and method

Results

Discussion

Conclusion

Abbreviations

Declarations

Author Contribution

References

Tables

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1