Abstract
Objectives
Metabolic dysfunction-associated steatotic liver disease (MASLD) and metabolic dysfunction-associated steatohepatitis (MASH) weakened the diagnostic ability of traditional elastography ultrasound on liver fibrosis. Combi-elasto ultrasound can simultaneously reflect the information of fibrosis degree and fat accumulation, thus have better potential and generality for liver fibrosis. We aim to develop a combi-elasto ultrasound based deep learning (DL) model for fibrosis staging in suspected MASLD/MASH patients.
Methods
Patients with overweight, type-2 diabetes, hypertension or dyslipidemia, and without alcoholic, drug-induced, and genetic hepatitis history were defined as suspected MASLD/MASH patients. Fibrosis stages were classified as none or mild (F0/F1), moderate (F2), and severe (F3/F4). Combi-elasto images from ten centers were collected to build the DL model. OpenCV was applied to automatically extract image information, Resnet-50 was used to analyze image feature, and adversarial training (AT) algorithm was introduced to enhance model generality. Traditional DL model without AT algorithm, and other seven classical fibrosis models were also built to compare with AT-DL model. Subgroup analysis on pathological confirmed normal, MASLD and MASH patients were performed to test model generality.
Results
From January 2022 to December 2023, 295 patients from eight centers were divided into training (n = 204) and validation (n = 91) sets. 101 patients from two independent centers were collected as external test set. In test set, AT-DL model achieved significantly higher accuracy than traditional DL model (0.871 vs. 0.812, p = 0.021) and all other classical models (0.545–0.762, all p < 0.05). For pathological confirmed normal patients, AT-DL and traditional DL models had comparable accuracy (0.881 vs. 0.905, p = 0.274). However, for pathological confirmed MASLD and MASH patients, AT-DL model showed significantly accuracy than traditional DL model (0.900 vs. 0.800, p = 0.028 and 0.828 vs. 0.690, p = 0.003).
Conclusion
Combi-elasto ultrasound based AT-DL model had satisfactory performance on fibrosis staging in suspected MASLD/MASH patients, and could reliable non-invasive monitoring of fibrosis progression.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02539-1.
Keywords: Liver fibrosis, Metabolic dysfunction-associated steatotic liver disease, Metabolic dysfunction-associated steatohepatitis, Combi-elasto, Deep learning
Summary statement
Combi-elasto based deep learning mode had satisfactory performance on fibrosis staging in suspected metabolic dysfunction-associated steatotic liver disease and metabolic dysfunction-associated steatohepatitis patients.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02539-1.
Key points
Combi-elasto based adversarial training deep learning model achieved satisfactory accuracy on fibrosis staging (87.1%, 95% CI: 81.2–92.1%) in suspected MASLD/MASH patients.
Compared with traditional deep learning model, adversarial training algorithm significantly improved accuracy in pathological confirmed MASLD (from 80.0% to 90.0%, p = 0.012) and MASH patients (from 69.0% to 82.8%, p = 0.003).
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02539-1.
Introduction
Global popular trend of metabolic dysfunction-associated steatotic liver disease (MASLD) and metabolic dysfunction-associated steatohepatitis (MASH) is intensifying in recent years [1, 2]. Previous studies have shown that MASLD and MASH can induce fibrosis, thereby increasing the risk of hepatocellular carcinoma (HCC) and portal hypertension (PHT) [3, 4]. In consequence, MASLD/MASH related HCC has increased 7.7 times in recent years and PHT also increased rapidly [5]. Therefore, accurate diagnosis and monitoring of fibrosis progression is essential to prevent MASLD/MASH patients from developing HCC and PHT [6, 7].
Elastography is the non-invasive method for liver fibrosis diagnosis recommended by many researches and guidelines [8–11]. Chen et al.‘s elastography model achieved 94% specificity on fibrosis staging for patients with chronic liver disease [8]. Dou et al.‘s elastography model also has an area under curve (AUC) above 0.80 for HBV patients [11]. However, elastography researches for MASLD or MASH patients were rare. There were two possible reasons: Firstly, elastography images had been interfered. MASLD/MASH-induced specific parenchymal alteration, such as diffuse fat infiltration, affected the propagation of mechanical waves, and reduced diagnostic performance [12]. Secondly, in-conformable to the clinical diagnostic process. Current diagnosis of MASLD/MASH depended on pathology, but once MASLD/MASH could be pathological-diagnosis, non-invasive diagnosis of fibrosis would be meaningless. Hisatoshi et al.‘s study established a model for fibrosis stage in MASLD patients with an AUC of 0.840 [13]. However, the inclusion criteria for this study were that patients must have a pathological-diagnosis of MASLD.
To overcome these difficulties, we adopted two new technologies: combi-elasto ultrasound and adversarial training (AT). Combi-elasto ultrasound was a comprehensive imaging examination that combined many traditional elastography information, including shear wave elastography (SWE) [14]and attenuation coefficient (ATT), which had been showed importance in the diagnosis of fibrosis and fatty liver [15]. AT was an intelligent method that could help deep learning (DL) model achieve better generality under uncertainty interference. For suspected MASLD/MASH patients without pathological diagnosis, ambiguous hepatic parenchymal status was the uncertainty interference for fibrosis assessment.
Therefore, our aim is to develop a combi-elasto ultrasound based model on fibrosis staging for suspected MASLD/MASH patients.
Method
Patients
This was a multicenter retrospective study. This study included patients from ten hospitals from January 2022 to December 2023. This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of Heping Hospital Affiliated to Changzhi Medical College (S-2022-14-5). Informed consent was obtained from all individual participants included in the study.
Suspected MASLD/MASH patient was defined as patient with overweight, type-2 diabetes, hypertension or dyslipidemia, and without alcoholic, drug-induced, and genetic hepatitis history. Therefore, the inclusion criteria were: (1) Age greater than 18 years; (2) Patients had at least one of the following: BMI ≥ 23, type-2 diabetes, hypertension or dyslipidemia; (3) Underwent biopsy within one month after combi-elasto ultrasound. The exclusion criteria were: (1) Excessive drinking history; (2) Drug-induced, or genetic hepatitis history; (3) Incomplete clinical information (Fig. 1). The clinical information included: sex, age, body mass index (BMI), type-2 diabetes history, hypertension history, viral hepatitis history, alanine transaminase (ALT), aspartate aminotransferase (AST), blood platelet (PLT), total protein (TP), and albumin (ALB). Based on the EASL clinical practice guidelines and AASLD practice guideline, we defined F0/1 as no or mild liver fibrosis, F2 as moderate fibrosis, and F3/4 as severe fibrosis [16, 17].
Fig. 1.
Flowchart. Patients from ten hospitals were collected and divided into training, validation, and test sets according to source hospital. All eligible patients had pathologically diagnosed liver fibrosis stages, and test set patients also had pathologically diagnosed liver background information, including normal, MASLD and MASH. BMI = Body mass index; MASLD = Metabolic dysfunction associated steatotic liver disease; MASH = Metabolic dysfunction associated steatohepatitis
Combi-elasto image and pathological tissue collection protocol
Combi-elasto images (ALOKAARIETTA, HITACHI, Japan) were collected at the area 1–2 cm below liver capsule and five images were collected at the same location. Parenchyma biopsy was as close as possible to the image location. All operators had at least five years of experience in abdominal ultrasound and received training in at least 50 cases of elastography.
Model construction
AT-DL model was composed by three modules: automatic annotation module, data augmentation module, and AT-DL module:
In automatic annotation module (Fig. 2a), OpenCV (Version 3.4 in Python 3.7) was applied to batch-process and accurately extract the region of interest (ROI) containing the combi-elasto area from original image. (Details in supplementary material).
Fig. 2.
Model establishment and experimental design. Deep learning model consisted of three modules: (a) Automatic annotation module, (b) Data augmentation module and (c) Adversarial deep learning module. (d) During model evaluation process, training, validation, and test sets used different processed datasets to complete the evaluation at the model level, image level, and patient level. ROI = Region of interest; ROC = Receiver operating characteristic; DL = Deep learning
In data augmentation module (Fig. 2b), ROIs were augmented by flipping, rotation, cropping, and translation through Tensorflow (Version 2.2 in Python 3.7). Data augmentation module was only used in training set (Details in supplementary material).
In AT-DL module (Fig. 2c), ResNet-50 (Tensorflow, version 2.2) was used as the DL model backbone, and AT algorithm was introduced to optimize network training (CleverHans, in Python 3.7). ResNet-50 was a classic DL algorithm had been used in many medical image studies. AT algorithm could intelligently explore the search space of attack strategies to generate the most threatening adversarial samples. Combined with the Gumbel Softmax sampling method, AT enabled efficient and stable adversarial training, ensuring the model maintained high robustness against various types of attacks (Details in supplementary material).
We also built a combi-elasto image based DL model without AT algorithm, a grayscale image based DL model (US model), a combi-elasto parameter model (Parameter model), as well as several classic fibrosis staging methods, including: SWE, ATT, Fibrosis-4 Index (FIB-4), aspartate aminotransferase-to-platelet ratio index (APRI), and NAFLD fibrosis score (NFS) (Details in supplementary material).
Model evaluation
Due to the data augmentation module, image collection protocol, and clinical practice conventions, we used different evaluation methods in training, validation, and test sets (Fig. 2d).
In training set, model evaluation was performed in the augmentation data set. In validation set, model predicted the fibrosis stage based on one image, so model evaluation was performed at image level. In test set, model predicted the fibrosis stage based on aggregating probabilities of five images from one patient, so model evaluation was performed at patient level.
Subgroup analyses on sex, age, BMI, and medical history were performed in test set. We also evaluated the AT-DL model performance on biopsy-confirmed normal, MASLD, and MASH patients, and compared its performance with traditional DL model to verify the anti-interference ability and generality from AT-DL module. Two senior radiologists were invited to classify patients into no or mild fatty liver, moderate fatty liver, and severe fatty liver based on ultrasound images (Details in supplementary material). Subsequently, we evaluated the AT-DL model performance in patients with different fatty liver degrees.
Statistical analysis
Continuous data were expressed as the means ± SDs and were evaluated by independent sample t-test or variance analysis. Categorical variables were analyzed using the χ2 test or Fisher’s exact test. The performance of the multi-classification model was displayed by the confusion matrix and macro-receiver operating characteristic (Macro-ROC) curve, and evaluated by accuracy, macro-area under curve (Macro-AUC), macro-specificity, macro-sensitivity, macro-precision, macro-recall, macro-F1 score, macro-positive predictive value (macro-PPV), and macro-negative predictive value (macro-NPV) (Details in supplementary material). The 95% confidence interval (CI) was evaluated by bootstrap. Interaction analysis was used to evaluate variation in model performance among different subgroups. Accuracy was compared by bootstrap, and p < 0.05 indicated a significant difference. Statistical analysis was performed using R (Version 4.3.2).
Results
From January 2022 to December 2023, a total of 396 eligible patients from ten hospitals were included in this study. Training set included 204 patients from six hospitals, validation set included 91 patients from two hospitals, the test set included 101 patients from two hospitals. There were no significant baseline differences among three sets (Table 1).
Table 1.
Baseline of training, validation and test sets
| Training set | Validation set | Test set | p value | ||
|---|---|---|---|---|---|
| 204 | 91 | 101 | |||
| Sex (%) | Women | 125 (61.3%) | 55 (60.4%) | 62 (61.4%) | 0.989 |
| Man | 79 (38.7%) | 36 (39.6%) | 39 (38.6%) | ||
| Age (year) | 42.35 ± 13.80 | 40.70 ± 12.42 | 41.52 ± 11.47 | 0.588 | |
| BMI | 25.64 ± 3.57 | 25.63 ± 3.64 | 25.48 ± 3.64 | 0.929 | |
| Viral hepatitis (%) | No | 124 (60.8%) | 56 (61.5%) | 56 (55.4%) | 0.611 |
| Yes | 80 (39.2%) | 35 (38.5%) | 45 (44.6%) | ||
| Diabetes (%) | No | 143 (70.1%) | 67 (73.6%) | 73 (72.3%) | 0.807 |
| Yes | 61 (29.9%) | 24 (26.4%) | 28 (27.7%) | ||
| Hypertension (%) | No | 145 (71.1%) | 59 (64.8%) | 71 (70.3%) | 0.548 |
| Yes | 59 (28.9%) | 32 (35.2%) | 30 (29.7%) | ||
| ALT (U/L) | 40.57 ± 25.05 | 39.24 ± 22.56 | 39.47 ± 27.06 | 0.889 | |
| AST (U/L) | 30.00 ± 17.41 | 29.51 ± 19.96 | 26.27 ± 17.47 | 0.223 | |
| PLT (10^9/L) | 119.67 ± 29.91 | 116.95 ± 26.72 | 120.26 ± 29.89 | 0.695 | |
| TP (g/L) | 71.87 ± 5.48 | 71.35 ± 5.54 | 71.19 ± 6.08 | 0.554 | |
| ALB (g/L) | 44.88 ± 4.10 | 44.82 ± 3.70 | 44.52 ± 3.86 | 0.754 | |
| Fibrosis staging | 0.334 | ||||
| None or mild | 108 (52.9%) | 48 (52.7%) | 52 (51.5%) | ||
| Moderate | 51 (25.0%) | 24 (26.4%) | 28 (27.7%) | ||
| Severe | 45 (22.1%) | 19 (20.9%) | 21 (20.8%) |
Note: BMI = Body Mass Index; ALT = Alanine transaminase; AST = Aspartate aminotransferase; PLT = Blood platelet; TP = Total protein; ALB = Albumin
Model construction
Automatic annotation module was designed based on computer vision technology rather than artificial intelligence technology, thus featuring exceptionally high operational speed (with a single image processing time of approximately 1 millisecond) and a compact size (module size under 20KB). Data augmentation module was used in training set and expand 1,020 images from 204 patients into 10,000 images.
Model evaluation
The training set contained 10,000 images. In training set, accuracy of AT-DL model was 0.926 (95%CI 0.882–0.966), Macro-AUC was 0.945 (95%CI 0.901–0.973), Macro-Specificity was 0.968 (95%CI 0.948–0.976), and Macro-Sensitivity was 0.927 (95%CI 0.888–0.967). Model performance of AT-DL model was comparable to traditional DL model (accuracy = 0.917 (95%CI 0.877–0.946), p = 0.623), and significantly higher than other models (US model, accuracy = 0.868 (95%CI 0.814–0.907); Parameter model, accuracy = 0.745 (95%CI 0.686–0.799); ATT, accuracy = 0.716 (95%CI 0.642–0.775), SWE, accuracy = 0.691 (95%CI 0.618–0.735), FIB-4, accuracy = 0.623 (95%CI 0.554–0.686), APRI, accuracy = 0.603 (95%CI 0.525–0.657), and NFS, accuracy = 0.598 (95%CI 0.534–0.667), all p < 0.05 (Table S1, Figure S1).
The validation set contained 455 combi-elasto images from 91 patients. In validation set, accuracy of AT-DL model was 0.890 (95%CI 0.813–0.945), Macro-AUC was 0.925 (95%CI 0.862–0.968), Macro-Specificity was 0.946 (95%CI 0.909–0.954), and Macro-Sensitivity was 0.899 (95%CI 0.813–0.954). Model performance of AT-DL model was comparable to traditional DL model (accuracy = 0.879 (95%CI 0.813–0.934), p = 0.547), and was significantly higher than other models (US model, accuracy = 0.824 (95%CI 0.747–0.890); Parameter model, accuracy = 0.681 (95%CI 0.582–0.791); ATT, accuracy = 0.659 (95%CI 0.560–0.736), SWE, accuracy = 0.637 (95%CI 0.527–0.725), FIB-4, accuracy = 0.593 (95%CI 0.495–0.670), APRI, accuracy = 0.582 (95%CI 0.484–0.681), and NFS, accuracy = 0.560 (95%CI 0.473–0.659), all p < 0.05) (Table S2, Figure S2).
The test set contained 101 patients, and the prediction of fibrosis stage was based on the averaging of predicted probabilities from five combi-elasto images provided by each patient. In test set, accuracy of AT-DL model was 0.871 (95%CI 0.812–0.931), Macro-AUC was 0.894 (95%CI 0.799–0.952), Macro-Specificity was 0.937 (95%CI 0.899–0.973), and Macro-Sensitivity was 0.873 (95%CI 0.808–0.926). Model performance of AT-DL model was significantly higher than that of traditional DL model (accuracy = 0.812 (95%CI 0.723–0.881), p = 0.021), and other models (US model, accuracy = 0.762 (95%CI 0.663–0.832); parameter model, accuracy = 0.624 (95%CI 0.525–0.693); ATT, accuracy = 0.624 (95%CI 0.515–0.713); SWE, accuracy = 0.604 (95%CI 0.465–0.693); FIB-4, accuracy = 0.574 (95%CI 0.485–0.663); APRI, accuracy = 0.545 (95%CI 0.446–0.644), and NFS, accuracy = 0.584 (95%CI 0.475–0.673), all p < 0.01) (Table 2; Fig. 3).
Table 2.
Model performance on liver fibrosis staging in test set
| AUC | 95% CI | ACC | 95% CI | SPE | 95% CI | SEN | 95% CI | |
|---|---|---|---|---|---|---|---|---|
| AT-DL model | 0.894 | (0.799–0.952) | 0.871 | (0.812–0.931) | 0.937 | (0.899–0.973) | 0.873 | (0.808–0.926) |
| Traditional DL model | 0.875 | (0.816–0.935) | 0.812 | (0.723–0.881) | 0.892 | (0.860–0.926) | 0.802 | (0.736–0.886) |
| US model | 0.819 | (0.741–0.885) | 0.762 | (0.663–0.832) | 0.891 | (0.827–0.916) | 0.76 | (0.653–0.834) |
| Parameter model | 0.71 | (0.596–0.781) | 0.624 | (0.525–0.693) | 0.826 | (0.767–0.846) | 0.626 | (0.527–0.696) |
| ATT | 0.743 | (0.676-0.800) | 0.624 | (0.515–0.713) | 0.784 | (0.755–0.826) | 0.627 | (0.527–0.713) |
| SWE | 0.693 | (0.595–0.762) | 0.604 | (0.465–0.693) | 0.815 | (0.725–0.853) | 0.607 | (0.478–0.704) |
| FIB-4 | 0.7 | (0.612–0.764) | 0.574 | (0.485–0.663) | 0.795 | (0.770–0.813) | 0.58 | (0.493–0.661) |
| APRI | 0.685 | (0.577–0.749) | 0.545 | (0.446–0.644) | 0.775 | (0.742–0.822) | 0.55 | (0.446–0.647) |
| NFS | 0.694 | (0.608–0.781) | 0.584 | (0.475–0.673) | 0.801 | (0.735–0.824) | 0.585 | (0.482–0.679) |
Note: AUC = Area Under Curve; ACC = Accuracy; SPE = Specificity; SEN =Sensitivity; CI = Confidence Interval; AT-DL model = Combi-elasto based adversarial training deep learning model; DL model = Combi-elasto based deep learning model; US = ultrasound; ATT = Attenuation coefficient; SWE = Shear wave elastography; FIB-4 = Fibrosis-4 index; APRI = Aspartate aminotransferase-to-platelet ratio index; NFS = NAFLD Fibrosis Score
Fig. 3.
Model performance on test set. AT-DL model = Combi-elasto based adversarial training deep learning model; US = ultrasound; ATT = Attenuation coefficient; SWE = Shear wave elastography; FIB-4 = Fibrosis-4 index; APRI = Aspartate aminotransferase-to-platelet ratio index; NFS = NAFLD Fibrosis Score
Subgroup analysis
In test set, we performed the subgroup analysis on different sex, age, BMI, disease history, pathological confirmed liver parenchyma (Normal, MASLD, and MASH) and imaging diagnosed fatty liver degrees (No or mild fatty liver, moderate fatty liver, and severe fatty liver). Fatty liver degrees were diagnosed by two radiologists (Kappa value = 0.77, Table S3). We also compared the model performed between AT-DL model and traditional DL model to assess the model’s generalizability and the effectiveness of AT algorithm.
For traditional DL model, there was no significant interaction between model performance with sex (p = 0.561), age (p = 0.439), BMI (p = 0.479), viral hepatitis history (p = 0.433), diabetes history (p = 0.377), and hypertension history (p = 0.843). However, significant interactions were existed between model performance with pathological confirmed liver parenchyma (p = 0.026) and imaging diagnosed fatty liver degrees (p = 0.011), which indicated that the performance of traditional DL model was interfered by these factors.
For AT-DL model, there was no significant interaction between model performance with sex (p = 0.551), age (p = 0.382), BMI (p = 0.881), viral hepatitis history (p = 0.472), diabetes history (p = 0.793), hypertension history (p = 0.928), pathological confirmed liver parenchyma (p = 0.550), and imaging diagnosed fatty liver degrees (p = 0.551), which indicated that the AT-DL model had very stable performance (Table 3).
Table 3.
Subgroup performance of traditional DL model and AT-DL model
| Traditional DL model | AT-DL model | p value | ||||
|---|---|---|---|---|---|---|
| ACC | Interaction | ACC | Interaction | |||
| Sex | 0.561 | 0.551 | ||||
| Women | n = 62 | 0.821 | 0.887 | |||
| Man | n = 39 | 0.806 | 0.846 | |||
| Age | 0.439 | 0.382 | ||||
| < 40 year | 0.838 | 0.897 | ||||
| >=40 year | 0.791 | 0.837 | ||||
| BMI | 0.479 | 0.881 | ||||
| < 23 | 0.84 | 0.88 | ||||
| >=23 | 0.803 | 0.868 | ||||
| Viral hepatitis | 0.433 | 0.472 | ||||
| No | 0.839 | 0.893 | ||||
| Yes | 0.778 | 0.844 | ||||
| Diabetes | 0.377 | 0.793 | ||||
| No | n = 73 | 0.821 | 0.877 | |||
| Yes | n = 28 | 0.786 | 0.857 | |||
| Hypertension | 0.843 | 0.928 | ||||
| No | n = 71 | 0.817 | 0.873 | |||
| Yes | n = 30 | 0.8 | 0.867 | |||
| Pathologic diagnosis | 0.026 | 0.55 | ||||
| Normal | n = 42 | 0.905 | 0.881 | 0.274 | ||
| MASLD | n = 30 | 0.8 | 0.9 | 0.028 | ||
| MASH | n = 29 | 0.69 | 0.828 | 0.003 | ||
| Radiologist assessment | 0.011 | 0.551 | ||||
| No or mild fatty liver | n = 50 | 0.9 | 0.9 | 0.999 | ||
| Moderate fatty liver | n = 29 | 0.793 | 0.828 | 0.456 | ||
| Severe fatty liver | n = 22 | 0.636 | 0.864 | > 0.001 | ||
Note: ACC = Accuracy; BMI = Body Mass Index; AT-DL model = Combi-elasto based adversarial training deep learning model; DL model = Combi-elasto based deep learning model; MASLD = Metabolic dysfunction associated steatotic liver disease; MASH = Metabolic dysfunction associated steatohepatitis
Interaction evaluated the interaction between model performance and subgroup
P value compared the performance between traditional DL model and AT-DL model
To elaborate, in pathological confirmed normal parenchyma group, performance of AI-DL model and traditional DL model on liver fibrosis staging was comparable (accuracy: 0.881 vs. 0.905, p = 0.354). In pathological confirmed MASLD group, AT-DL model achieved significantly higher accuracy than that of traditional DL model (0.900 vs. 0.800, p = 0.012). In pathological confirmed MASH group, AT-DL model also outperformed traditional DL model significantly (0.828 vs. 0.690, p = 0.003) (Table 3; Fig. 4a).
Fig. 4.
Performance of AT-DL model in pathologically diagnosed normal, MASLD, and MASH patients. (a) Confusion matrix showed that AT-DL model had satisfactory and stable performance in normal, MASLD and MASH patients; (b) Multiple indices of AT-DL model performance in normal, MASLD and MASH patients were all over 0.80. AT-DL model = Combi-elasto based adversarial training deep learning model; MASLD = Metabolic dysfunction associated steatotic liver disease; MASH = Metabolic dysfunction associated steatohepatitis; AUC = Area under curve; ACC = Accuracy; SPE = Specificity; SEN = Sensitivity; PPV = Positive predictive value; NPV = Negative predictive value
In imaging diagnosed no or mild fatty liver group, performance of AT-DL model and traditional DL model was comparable (0.900 vs. 0.900, p = 0.999). In imaging diagnosed moderate fatty liver group, performance of AT-DL model and traditional DL model was comparable (0.828 vs. 0.793, p = 0.214). In imaging diagnosed severe fatty liver group, AT-DL model demonstrated significantly higher accuracy compared to traditional DL model (0.864 vs. 0.636, p < 0.001) (Table 3 and Figure S4a).
Finally, we comprehensively evaluated AT-DL model performance in different pathological confirmed liver parenchyma and imaging diagnosed fatty liver degrees. The Accuracy, Macro-AUC, Macro-Specificity, Macro-Sensitivity, Macro-precision, Macro-recall, Macro-F1, Macro-PPV, and Macro-NPV of CE-AT model were over 0.80 in all subgroups (Fig. 3b and Figure S4b).
Discussion
In this study, we developed a combi-elasto ultrasound based adversarial training deep learning model for liver fibrosis staging in suspected MASLD/MASH patients. This model achieved well diagnostic performance (accuracy = 0.871) and satisfactory generality, which was significantly higher than current clinical methods (all p < 0.05). Our model solved a common but troublesome clinical problem: when doctors encountered suspected MASLD/MASH patients, the diagnostic result of elastography for fibrosis staging would be less credible. And patients had to recourse to other imaging examination, such as MRI, or invasive histological assessment, inevitably increasing healthcare costs and patient discomfort. In this study, AT-DL model significantly enhanced diagnostic reliability and accuracy in suspected MASLD/MASH patients, enabling doctors to accurately and confidently perform non-invasive diagnosis of fibrosis staging in a cost-effective manner [18].
With the escalating global burden of MASLD and MASH, incidence of associated HCC and PHT was increasing rapidly [5, 6]. Previous studies proposed two mechanisms by which MASLD and MASH lead to HCC and PHT [19, 20]. First, the indirect mechanism, which meant that MASLD/MASH could induce fibrosis progression, and fibrosis subsequently elevated the risk of HCC and PHT [3, 21]. Second, the direct mechanism, which meant MASLD/MASH could directly lead to HCC/PHT, and the synergistic effect between MASLD/MASH with fibrosis would enhance this process [22, 23]. In both two mechanisms, fibrosis played the key role in disease progression [24, 25]. Therefore, in the context of increasing prevalence of MASLD and MASH, establishing a non-invasive monitoring method on fibrosis progress that could be widely applied to suspected MASLD/MASH patients was of great significance for reducing the risk of PHT [26] and improving the success rate of ablation for HCC [27].
Conventional fibrosis staging models were constrained by specific populations [9, 28, 29], such as Chen et al.‘s model in patients with chronic liver disease and Dou et al.‘s model in patients with hepatitis B [8, 30, 31]. However, this constraint significantly impacted the clinical application of models for MASLD patients or MASH patients. Although Hisatoshi et al.‘s model achieved an AUC of 0.840 for fibrosis staging in MASLD patients, it also required that patients eligible for this model must have a pathological confirmed MASLD diagnosis. This constraint might also be the crucial factor led to the absence of models for fibrosis staging in patients with MASLD or MASH.
Quantitative Ultrasound (QUS) was a common diagnostic method for steatosis, which could provide quantitative parameters such as attenuation coefficient and backscatter coefficient. However, studies using QUS alone for fibrosis staging were very rare, likely due to pathophysiological limitations. The primary change in fibrosis was increased tissue stiffness, whereas the attenuation coefficient mainly reflected fat content, and the backscatter coefficient was influenced by collagen deposition. In recent years, some studies developed machine learning models for fibrosis staging by integrating QUS parameters with elastography parameters. Tang et al. developed a random forest model combining attenuation coefficient and liver stiffness [32]. This model achieved AUCs of 0.72, 0.77, 0.77, and 0.75 for classification of fibrosis stages F0 vs. F1–4, F0–1 vs. F2–4, F0–2 vs. F3–4, and F0–3 vs. F4, respectively. Tsui et al. collected multimodal QUS parameters from patients with fatty liver disease and built machine learning models to distinguish early fibrosis (≥ F1, AUC: 0.78–0.81) and significant fibrosis (≥ F2, AUC: 0.64–0.76) [33]. Nevertheless, even with the combination of QUS and elastography parameters and focusing only on binary classification, the performance of these models remained insufficient and unstable, falling below that of our model (accuracy = 0.871). This might be because QUS and elastography parameters were simplifications of the original ultrasound images and did not fully capture the image information. In contrast, deep learning model had a more fundamental information source, as they could automatically learn task-relevant features directly from the data.
In our study, adversarial training enabled AT-DL model to achieve effective and stable performance in suspected MASLD/MASH patients through analyzing images from relaxed inclusion criteria, so the relaxed inclusion criteria allowed model to apply in wider population. In fact, with the increasing integration between macro-information and micro-information, studies similar to predicting fibrosis staging in suspected MASLD/MASH patients through imaging would become more prevalent. For instance, in HCC patients, some studies aimed to build imaging-based model to predict microvascular invasion in vascular endothelial growth factor positive patients. For these studies, adversarial training could provide a novel method to build unrestricted models for real clinical application.
In various imaging examinations, elastography was a cost-effective, radiation-free technology, particularly suited for medical screening in the expanding suspected MASLD/MASH patients, because the indiscriminate CT/MRI utilization for those patients might lead to the healthcare overexpenditure and overdiagnosis. However, the application of elastography diagnosis in suspected MASLD/MASH patients faced two critical limitations: Firstly, traditional elastography oversimplified biomechanics by focusing solely on tissue elasticity while neglecting viscoelastic parameters crucial for MASLD and MASH associated fibrotic remodeling [14]; Secondly, the pathophysiological interplay between steatosis induced acoustic impedance changes and inflammatory extracellular matrix alterations created confounding “noise” that traditional DL architectures failed to disentangle [34].
To address these limitations, we adopted two corresponding improvements. Firstly, the innovative combi-elasto ultrasound technique was developed to deliberately integrate and leverage the distinct technical advantages offered by conventional shear wave elastography, which provided quantitative stiffness measurements, and real-time tissue elastography, which offered qualitative visual assessment of tissue deformation patterns; this strategic fusion was crucial as it enabled a significantly more comprehensive, multi-dimensional, and nuanced reflection of the underlying parenchyma information, capturing not just stiffness but also deformation characteristics across different tissue planes, thereby establishing a robust and information-rich foundation essential for building highly accurate and generalizable diagnostic models [35, 36]. Secondly, we incorporated an AT module, a sophisticated mechanism engineered to proactively generate carefully crafted adversarial samples that simulated plausible real-world input distribution deviations and challenging edge cases the model might encounter post-deployment; this process actively forced the model to learn more robust feature representations and decision boundaries, substantially enhancing its stability, reliability, and predictive accuracy when confronted with previously unseen, inherently noisy, or subtly corrupted clinical data inputs, while a key intelligent aspect of this module was its ability to dynamically and adaptively calibrate the intensity and nature of the adversarial perturbations it introduced throughout the training epochs, meticulously adjusting the challenge level based on the model’s current learning status and performance metrics to ensure progressive hardening without causing the model to detrimentally overfit or become overly sensitive to any specific, narrow type of adversarial sample, thus promoting superior generalization capabilities critical for clinical utility [37, 38]. Through the above two improvements, our model achieved excellent and stable performance for fibrosis staging in suspected MASLD/MASH patients. Our model also integrated the automatic annotation module, which had sophisticated and fast ROI extraction capabilities, making the application of the model more feasible. In addition, our comparison between AT-DL and traditional DL demonstrated the value of AT module. In the broader context of domain generalization in ultrasound AI, other strategies such as style transfer, domain‑adversarial neural networks, and meta‑learning had also been explored to address cross‑center or cross‑device variability. While our AT approach focused on robustness to input‑space perturbations, these feature‑space alignment methods were complementary. In future work, a combination may further enhance generalizability.
Our model achieved an accuracy of more than 0.82 in patients with different genders, ages, BMI, and medical histories. For MASLD patients with hepatitis B and MASH patients with hepatitis B, the accuracy of our model was 0.818 and 0.833, respectively. For lean MASLD patients and lean MASH patients (BMI less than 23), the accuracy of our model was 0.800 and 0.857, respectively. Furthermore, our multicenter design provided strong evidence of generalizability across different clinical settings and patient populations, reducing the risk of overfitting to a single center’s practice. This result demonstrated that the performance of our model was not unaffected by complex clinical conditions.
There were several limitations in our study. First, this study was a retrospective study, and further testing of our model needed to be conducted in more complex prospective data. Second, this study only included a single race group from a single country, and the lack of racial diversity still affected model robustness. Third, since combi-elasto ultrasound was not yet widely available, the use of a single ultrasound device in this study might affect model generalizability. Future work would attempt to evaluate the model’s performance on other devices with similar functions.
Conclusion
The combi-elasto ultrasound based adversarial training deep learning model achieved superior diagnostic performance and satisfactory generality for liver fibrosis staging. This model provided a clinically viable solution for non-invasive longitudinal monitoring of fibrotic progression in suspected MASLD/MASH patients.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
No.
Abbreviations
- MASLD
Metabolic Dysfunction-Associated Steatotic Liver Disease
- MASH
Metabolic Dysfunction-Associated Steatohepatitis
- HCC
Hepatocellular Carcinoma
- PHT
Portal Hypertension
- ALT
Alanine transaminase
- AST
Aspartate aminotransferase
- PLT
Blood platelet
- TP
Total protein
- ALB
Albumin
- AI
Artificial Intelligence
- DL
Deep Learning
- ROI
Region of Interest
- AT
Automated Adversarial Training
- AT-DL model
Combi-elasto based adversarial training deep learning model
- ATT
Attenuation coefficient
- SWE
Shear Wave Elastography
- FIB-4
Fibrosis-4 Index
- APRI
Aspartate Aminotransferase-to-Platelet Ratio Index
- NFS
NAFLD Fibrosis Score
- CI
Confidence Interval
Author contributions
XDL, PF and WXY conceived and designed the study. YMH and DL acquired the data. LS and ZQP did the statistical analyses. XDL developed, trained, and applied the artificial neural network. LX verified the underlying raw data. All authors had access to the data presented in the manuscript. All authors analysed and interpreted the data. XDL prepared the first draft of the manuscript. LX and WXY revised the manuscript. All authors contributed to manuscript preparation. All authors were responsible for the decision to submit the manuscript for publication.
Funding
No.
Data availability
Authors will share deidentified individual participant imaging data on request with researchers who provide a methodologically viable proposal and do analyses that achieve the aims of the proposal. Data sharing requests can be directed to XDL by email.
Declarations
Ethics approval and consent to participate
This study was approved by the Ethics Committee of Heping Hospital Affiliated to Changzhi Medical College (S-2022-14-5). Written informed consent was obtained from all individual participants included in the study, after they had been fully informed of the study purpose, methods, potential risks, and benefits. The informed consent process complied with the Declaration of Helsinki and relevant national regulations. All participants voluntarily agreed to participate and were aware that their data would be used for scientific analysis.
Consent for publication
Not applicable. This manuscript does not contain any individual person’s data, images, or clinical details that would compromise anonymity, and therefore no specific consent for publication is required.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Contributor Information
Xin Li, Email: lixin301@hotmail.com.
Xiaoyan Wang, Email: wangxiaoyan@czmc.edu.cn.
References
- 1.Zobair MY, Homie R, Michael S, Alina MA, Quentin MA, Kenneth C, Scott LF, Eric L, Jeffrey VL, Detlef S, et al. Addressing the high and rising global burden of metabolic dysfunction-associated steatotic liver disease (MASLD) and metabolic dysfunction-associated steatohepatitis (MASH): From the growing prevalence to payors’ perspective. Aliment Pharmacol Ther. 2025;61(9). [DOI] [PubMed]
- 2.Scott LF. Fat, fibrosis, and the future: navigating the maze of MASLD/MASH. J Clin Invest. 2025;135(7). [DOI] [PMC free article] [PubMed]
- 3.Huang DQ, Singal AG, Kanwal F, Lampertico P, Buti M, Sirlin CB, Nguyen MH, Loomba R. Hepatocellular carcinoma surveillance — utilization, barriers and the impact of changing aetiology. Nat Reviews Gastroenterol Hepatol. 2023;20(12):797–809. [DOI] [PubMed] [Google Scholar]
- 4.Lin E, Zou B, Zeng G, Cai C, Li P, Chen J, Li D, Zhang B, Li J. The impact of liver fibrosis on microvascular invasion and prognosis of hepatocellular carcinoma with a solitary nodule: a Surveillance, Epidemiology, and End Results (SEER) database analysis. Ann Transl Med. 2021;9(16):1310. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Huang DQ, El-Serag HB, Loomba R. Global epidemiology of NAFLD-related HCC: trends, predictions, risk factors and prevention. Nat Reviews Gastroenterol Hepatol. 2020;18(4):223–38. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Cholankeril G, Kramer JR, Chu J, Yu X, Balakrishnan M, Li L, El-Serag HB, Kanwal F. Longitudinal changes in fibrosis markers are associated with risk of cirrhosis and hepatocellular carcinoma in non-alcoholic fatty liver disease. J Hepatol. 2023;78(3):493–500. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.David EK, Elizabeth MB, Mark VN, Cynthia B, Melissa JC, Oscar WC, Linda DF, Yao-Chang L, Michael ST, Aynur U-A, et al. Design and validation of a histological scoring system for nonalcoholic fatty liver disease. Hepatology. 41(6). [DOI] [PubMed]
- 8.Chen LD, Huang ZR, Yang H, Cheng MQ, Hu HT, Lu XZ, Li MD, Lu RF, He DN, Lin P, et al. US-based Sequential Algorithm Integrating an AI Model for Advanced Liver Fibrosis Screening. Radiology. 2024;311(1):e231461. [DOI] [PubMed] [Google Scholar]
- 9.Hassoun S, Bruckmann C, Ciardullo S, Perseghin G, Marra F, Curto A, Arena U, Broccolo F, Di Gaudio F. NAIF: A novel artificial intelligence-based tool for accurate diagnosis of stage F3/F4 liver fibrosis in the general adult population, validated with three external datasets. Int J Med Informatics. 2024;185. [DOI] [PubMed]
- 10.Rui F, Xu L, Yeo YH, Xu Y, Ni W, Tan Y, Zheng Q, Tian X, Zeng Q-L, He Z, et al. Machine Learning–Based Models for Advanced Fibrosis and Cirrhosis Diagnosis in Chronic Hepatitis B Patients With Hepatic Steatosis. Clin Gastroenterol Hepatol. 2024;22(11):2250–e22602212. [DOI] [PubMed] [Google Scholar]
- 11.Dou J, Chen X, Cheng G, Meng F, Zheng Y, Zhang B, Zhang W, Zhang Y, Yin Z, Yang H, et al. Dual elastography to discriminate adjacent stages of fibrosis and inflammation in chronic hepatitis B: A prospective multicenter study. Hepatology. 2023;79(2). [DOI] [PubMed]
- 12.Hammerich L. Tacke FJNrG, hepatology: Hepatic inflammatory responses in liver fibrosis. 2023. [DOI] [PubMed]
- 13.Hisatoshi H, Goki S, Masatsugu O, Koji O, Nobuharu T, Hayato H, Hiroaki H, Shinya M, Masaya S, Tatsuhiko K, et al. Elevated A2F bisect N-glycans of serum IgA reflect progression of liver fibrosis in patients with MASLD. J Gastroenterol. 2025;60(4). [DOI] [PMC free article] [PubMed]
- 14.Madalina-Gabriela I, Dan-Corneliu L, Monica L-P, Laura T, Silvia F, Ahmed H, Olga Hilda O, Bogdan P, Horia S, Maria Cristina M, et al. Diagnostic accuracy of 2D-SWE ultrasound for liver fibrosis assessment in MASLD: A multilevel random effects model meta-analysis. Hepatology. 2024(0). [DOI] [PMC free article] [PubMed]
- 15.Sheka AC, Adeyi O, Thompson J, Hameed B, Crawford PA, Ikramuddin S. Nonalcoholic steatohepatitis. JAMA. 2020;323(12). [DOI] [PubMed]
- 16.Berzigotti A, Tsochatzis E, Boursier J, Castera L, Cazzagon N, Friedrich-Rust M, Petta S, Thiele M. EASL Clinical Practice Guidelines on non-invasive tests for evaluation of liver disease severity and prognosis – 2021 update. J Hepatol. 2021;75(3):659–89. [DOI] [PubMed] [Google Scholar]
- 17.Richard KS, Andres D-R, Keyur P, Sumeet KA, Mouaz A, Jonathan AD, Maria Isabel F, Hassan MM, Daniel HL, Deborah L. AASLD practice guideline on imaging-based noninvasive liver disease assessment of hepatic fibrosis and steatosis. Hepatology. 2024;81(2). [DOI] [PubMed]
- 18.Bryan AP-P, Rocío G-D, Berenice MR-C, José Antonio V-RV, Manuel R-G, Jordi G-S. Advancing precision medicine in metabolic dysfunction-associated steatotic liver disease. Trends Endocrinol Metab. 2025(0). [DOI] [PubMed]
- 19.Zhang R, Chen J, Jiang Y, Wang J, Chen S. Prognostic nomogram for hepatocellular carcinoma with fibrosis of varying degrees: a retrospective cohort study. Ann Transl Med. 2020;8(21):1429. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Argenziano ME, Kim MN, Montori M, Di Bucchianico A, Balducci D, Ahn SH, Svegliati Baroni G. Epidemiology, pathophysiology and clinical aspects of Hepatocellular Carcinoma in MAFLD patients. Hep Intl. 2024;18(S2):922–40. [DOI] [PubMed] [Google Scholar]
- 21.Harrison SA, Allen AM, Dubourg J, Noureddin M, Alkhouri N. Challenges and opportunities in NASH drug development. Nat Med. 2023;29(3):562–73. [DOI] [PubMed] [Google Scholar]
- 22.Hanamatsu H, Suda G, Ohara M, Ogawa K, Tamaki N, Hikita H, Haga H, Maekawa S, Sugiyama M, Kakisaka T, et al. Elevated A2F bisect N-glycans of serum IgA reflect progression of liver fibrosis in patients with MASLD. J Gastroenterol. 2025. [DOI] [PMC free article] [PubMed]
- 23.Loomba R, Friedman SL, Shulman GI. Mechanisms and disease consequences of nonalcoholic fatty liver disease. Cell. 2021;184(10):2537–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Ezeani C, Omaliko C, Al-Ajlouni YA, Njei B. Mortality, hepatic decompensation, and cardiovascular- and renal-related outcomes in lean versus non-lean patients hospitalized with metabolic dysfunction-associated steatohepatitis (MASH). Cureus. 2024. [DOI] [PMC free article] [PubMed]
- 25.Ji Hye J, Kuo D, Rajesh Kumar D, Raquel M-D, Seh Hoon O, Liuyang W, Guannan G, Ana F, Jon H, Steven SP, et al. The senescence-associated secretome of hedgehog-deficient hepatocytes drives MASLD progression. J Clin Invest. 2024;134(19). [DOI] [PMC free article] [PubMed]
- 26.Liu S, Jiang X, Fu J, Wong VW-S, Zhong VW, Qi X. Baseline and Dynamic MAF-5 Score to Predict Liver Fibrosis and Liver-Related Events in General Population With MASLD. Clin Gastroenterol Hepatol. 2025;23(2):365–e368363. [DOI] [PubMed] [Google Scholar]
- 27.Ding W, Zheng L, Wei H, Li X, Jiang T, Xue G, Lu M, Cheng Z, Han Z, Yu X, et al. Artificial intelligence system assists doctors to improve the treatment effect of microwave ablation on hepatocellular carcinoma: A prospective cohort study. Digit Eng. 2025;6.
- 28.Carnazzo V, Pignalosa S, Tagliaferro M, Gragnani L, Zignego AL, Racco C, Di Biase L, Basile V, Rapaccini GL, Di Santo R, et al. Exploratory study of extracellular matrix biomarkers for non-invasive liver fibrosis staging: A machine learning approach with XGBoost and explainable AI. Clin Biochem. 2025;135. [DOI] [PubMed]
- 29.Ding W, Wu J, Dou J, Liu F, Han Z, Yu J, Liang P. Biologically explicable multimodal model predicting local tumor progression and tumor invasiveness of hepatocellular carcinoma. Hepatology. 2026. [DOI] [PubMed]
- 30.Jianping D, Xi C, Guangwen C, Fankun M, Ying Z, Bulin Z, Wei Z, Yao Z, Zhiyong Y, Hong Y, et al. Dual elastography to discriminate adjacent stages of fibrosis and inflammation in chronic hepatitis B: A prospective multicenter study. Hepatology. 2024;79(2):438–50. [DOI] [PubMed] [Google Scholar]
- 31.Ding W, Meng Y, Ma J, Pang C, Wu J, Tian J, Yu J, Liang P, Wang K. Contrast-enhanced ultrasound-based AI model for multi-classification of focal liver lesions. J Hepatol. 2025. [DOI] [PubMed]
- 32.Destrempes F, Gesnik M, Chayer B, Roy-Cardinal M-H, Olivié D, Giard J-M, Sebastiani G, Nguyen BN, Cloutier G, Tang A. Quantitative ultrasound, elastography, and machine learning for assessment of steatosis, inflammation, and fibrosis in chronic liver disease. PLoS ONE. 2022;17(1):e0262291. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Chuang Y-W, Le DC, Wang C-Y, Tai D-I, Zhou Z, Tsui P-H. Early Detection of Liver Fibrosis Using Scatteromics Based on Multimodal QUS Envelope Statistics Imaging. Diagnostics (Basel). 2026;16(4):564. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Thomas JW, Pichamol J, Raj S, Kimberly S, David JP, Christopher CT, Laura D, David BL, Marvin R. EUS-guided shear wave elastography for fibrosis screening in patients with obesity and metabolic dysfunction-associated steatotic liver disease: a pilot study (with video). Gastrointest Endosc. 2024;101(2). [DOI] [PubMed]
- 35.Paul YK, Howard CM, Esperance AS, Lawrence SF. Evaluation of abnormal liver biochemical test results. Gastroenterology. 2026(0). [DOI] [PubMed]
- 36.Claudia MV, Diana vdV, Miguel AQ-S, David FR. A strategic guide of techniques for biomedical and tissue engineering applications to measure mechanical properties of soft matter, eye and skin. Adv Drug Deliv Rev. 2026;231(0). [DOI] [PubMed]
- 37.Qin R, Wang L, Du X, Xie P, Chen X, Yan B. Adversarial robustness in deep neural networks based on variable attributes of the stochastic ensemble model. Front Neurorobot. 2023;17:1205370. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Guo F, Sun Z, Chen Y, Ju L. Towards the universal defense for query-based audio adversarial attacks on speech recognition system. Cybersecurity. 2023;6(1).
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Authors will share deidentified individual participant imaging data on request with researchers who provide a methodologically viable proposal and do analyses that achieve the aims of the proposal. Data sharing requests can be directed to XDL by email.