Computed tomography-based transformer model for non-invasive prediction of portal venous pressure in chronic liver diseases

Abstract

BACKGROUND

The hepatic venous pressure gradient (HVPG) is critical for the assessment of portal hypertension, but measuring the HVPG is invasive and costly. Current radiomics machine learning prediction models still exhibit certain limitations in clinical practice.

AIM

To investigate a new transformer-based regression model for predicting the HVPG in patients with chronic liver diseases non-invasively.

METHODS

This retrospective study randomly divided eligible patients from September 2017 to August 2023 into the training, internal testing and temporally independent validation sets. Computed tomography images from four sections were pre-processed and then input into the transformer network combined with transfer learning strategy for feature extraction, regression analysis and prediction of the HVPG values. The prediction and diagnostic performance of our model was investigated and compared with that of other non-invasive prediction methods across distinct HVPG thresholds.

RESULTS

The original datasets included 228 patients for training, 57 for internal testing, and 74 for temporally independent validation sets. The predictive HVPGs were strongly correlated with (r = 0.949, 0.741 and 0.717, P < 0.001) and showed acceptable agreement with the actual HVPGs in all the datasets. The overall prediction accuracies of our model for binary, ternary, and quinary HVPG classification intervals were 90.8%, 80.4% and 64.0%, respectively. For different binary classification intervals (HVPG > 5, ≥ 10, ≥ 12, ≥ 16, and ≥ 20 mmHg), the majority of area under the receiver operating characteristic curve in the three datasets for our model prediction of the HVPG exceeded 0.90, which outperformed conventional non-invasive HVPG prediction methods.

CONCLUSION

A new regression machine learning model based on the transformer network has been developed to be promising for non-invasively predicting the HVPG across the entire spectrum of portal venous pressure.

Key Words: Portal hypertension; Hepatic venous pressure gradient; Transformer network; Computed tomography; Non-invasive prediction

Core Tip: This retrospective study developed and validated a non-invasive regression machine learning model to predict the hepatic venous pressure gradient values in chronic liver diseases. Utilizing the transformer network combined with transfer learning based on contrast-enhanced computed tomography images from four pivotal sections, our model efficiently predicted the entire spectrum of portal venous pressure (from 1 to 37 mmHg) and showed good diagnostic performance for clinically significant hepatic venous pressure gradient thresholds including 5, 10, 12, 16 and 20 mmHg, outperforming conventional non-invasive methods. Moreover, the time-saving approach in this study minimized region of interest delineation workload, demonstrating potential for routine clinical application.

INTRODUCTION

Portal hypertension develops with the progression of chronic liver diseases and is recognized as the most important pathophysiological alteration in patients with liver cirrhosis[1]. Significant portal hypertension can incite severe or even fatal decompensation events, such as refractory ascites and esophagogastric variceal bleeding (EGVB), worsening the prognosis of patients[2]. Therefore, the value of portal venous pressure provides physicians with critical information regarding patients with liver diseases.

The hepatic venous pressure gradient (HVPG) is strongly associated with clinical endpoints in patients with portal hypertension[3] and is currently considered the gold standard for quantifying portal venous pressure[4]. However, transjugular HVPG measurement is an invasive and costly procedure requiring specialized personnel and equipment. Thus, the HVPG is not accessible for routine clinical practice or repeated measurements during the long-term follow-up.

In recent years, numerous non-invasive approaches, including liver stiffness measurement (LSM) and serum-based indices, have been investigated for the assessment of portal hypertension, but their prediction performance remains insufficient for widespread clinical application[5,6]. Advances in machine learning have facilitated the development of image-based models for non-invasive portal venous pressure evaluation, although several limitations persist. For example, most existing models based on radiomics or convolutional neural network (CNN) were primarily designed to identify clinically significant portal hypertension (CSPH) through binary classification strategies[7,8]. While these models achieved promising diagnostic performance, they generally provided limited information regarding the entire spectrum of portal venous pressure, especially for patients with normal (HVPG ≤ 5 mmHg) and mild portal hypertension (5 mmHg < HVPG < 10 mmHg), which may represent the early stages of chronic liver diseases. Moreover, some previously reported virtual HVPG models required extensive image pre-processing, with approximately 2-3 hours needed for each patient[9].

Unlike CNN, the transformer network is a seldomly used network of machine learning in portal venous pressure prediction. It can capture complex relationships between all the elements in a sequence via self-attention mechanisms to process image patches[10]. The principle of the transformer network allows the tackling of considerable radiological images with a comprehensive view and has gained widespread adoption in medical image classification tasks[11]. The advantages of the transformer network might help to overcome the shortcomings of current machine learning models in predicting the HVPG. Therefore, our study aimed to establish a non-invasive regression model with the transformer network based on computed tomography (CT) images to predict continuous HVPG values across various degree of portal venous pressure in patients with chronic liver diseases.

MATERIALS AND METHODS

Study design and patients

The study was a retrospective, single-center study conducted at West China Hospital, Sichuan University, and it was approved by the Biomedical Ethics Committee of West China Hospital, No. 20211488. Informed consent was waived. Patients with chronic liver diseases from September 2017 to August 2023 were included in this study if they met all the inclusion criteria and did not meet any exclusion criteria. The inclusion criteria were as follows: (1) Aged 18-85 years; (2) Underwent the HVPG measurement; and (3) Had access to imaging via abdominal contrast-enhanced CT within 14 days prior to the HVPG measurement. The exclusion criteria were as follows: (1) Non-cirrhotic portal hypertension; (2) Cavernous transformation of the portal vein; (3) History of transjugular intrahepatic portosystemic shunt or surgical shunt; (4) History of liver surgery, including resection and transplantation; (5) History of splenectomy or splenic artery embolization; and (6) Liver malignancies.

The included patients from September 2017 and July 2021 were randomly divided into the training set and the internal testing set at a ratio of 4:1 through a five-fold cross validation. The included patients from June 2022 and August 2023 were assigned to the temporally independent validation set. Clinical data, including gender, age, etiology of chronic liver diseases, clinical complications, Child-Pugh classification, endoscopic therapy and the measured HVPG values, were collected. In addition, data from laboratory tests, including platelet counts, total bilirubin, aspartate aminotransferase, alanine aminotransferase, gamma-glutamyl transferase, alkaline phosphatase, albumin, globulin, blood ammonia, prothrombin time, and international normalized ratio, were obtained within 7 days before the HVPG measurement.

HVPG measurement and classification intervals

The process of HVPG measurement could be seen in Supplementary material. HVPG > 5 mmHg is widely recognized as the earliest threshold indicating the presence of portal hypertension by domestic and international guidelines. 5 mmHg < HVPG < 10 mmHg is considered a subclinical stage of portal hypertension, reflecting early sinusoidal and microvascular alterations prior to overt clinical complications. HVPG ≥ 10 mmHg is defined as CSPH, which is associated with decompensation, while HVPG ≥ 12, ≥ 16, and ≥ 20 mmHg respectively indicate high risk of EGVB, increased death rate, and uncontrolled gastrointestinal hemorrhage, rebleeding or mortality[1,2,12]. On the basis of the HVPG thresholds with established clinical significance mentioned above, we set three types of HVPG classification intervals to test the granularity of our machine learning model as follows: (1) Binary: HVPG ≤ 5 mmHg and HVPG > 5 mmHg; (2) Ternary: HVPG ≤ 5 mmHg, 5 mmHg < HVPG < 12 mmHg, and HVPG ≥ 12 mmHg; and (3) Quinary: HVPG ≤ 5 mmHg, 5 mmHg < HVPG < 12 mmHg, 12 mmHg ≤ HVPG < 16 mmHg, 16 mmHg ≤ HVPG < 20 mmHg, and HVPG ≥ 20 mmHg.

CT image acquisition and pre-processing

In this study, upper abdominal contrast-enhanced CT was performed in accordance with the standard protocol (Supplementary material). Due to the architectural requirement of the transformer network, all the CT images were exported and saved in Joint Photographic Experts Group format through the Siemens Intelligent Imaging System with RGB color encoding. Each image had a resolution of 1636 × 1006 pixels, with a file size of approximately 90-150 kB.

Image pre-processing included region of interests (ROIs) delineation, image normalization, and image rearrangement. The delineation of ROIs was carried out utilizing the image annotation software “LabelMe” (Supplementary material). We chose four horizonal images in the portal venous phase, including the first hepatic hilum section (section a), the second hepatic hilum section (section b), the section of the splenic vein merging into the portal vein (section c), and the section with the largest spleen area (section d). Images of different sections did not share the same ROIs. In detail, the ROIs in section a were the portal vein, liver, and spleen, and the ROIs in section b were the hepatic vein, liver, and spleen. In addition, the ROIs of section c were the splenic vein, portal vein, and spleen, whereas the ROIs in section d were the spleen. On the basis of the above labeled ROIs, their pixels were extracted for superimposition and were retained, whereas the pixels of unlabeled ROIs were assigned a value of zero to achieve image segmentation. For image normalization, we adopted the average CT value of the aorta abdominalis to balance the luminance of all the images and histogram equalization to enhance image quality. The variation in the number of CT image slices across the four anatomical levels for each patient was addressed by implementing a data rearrangement strategy after dataset splitting, thus augmenting the quantity of image data through permuting and combining the CT images. All rearrangement images generated from the same patient remained within the same dataset during the development and evaluation of the model.

Establishment of the HVPG prediction model

The process of HVPG prediction model construction was illustrated in Figure 1. All experiments were conducted on NVIDIA Tesla A800 GPUs. After standardized pre-processing, each original CT image was resized to 224 × 224 pixels and then divided into 16 non-overlapping patches of 14 × 14 pixels. Each image patch was transformed into a fixed-length vector via a linear embedding layer. To retain spatial structure, positional encoding was added to each patch embedding, and a trainable prompt token was appended to guide the model’s learning. Then the embedded vectors, along with a prepended class token, were organized into an independent sequence and input into the transformer-based backbone network for feature extraction. In this study, we adopted the visual prompt tuning framework, where the transformer parameters were frozen and only the prompt tokens and task-specific heads were trainable. On this basis, the resulting image-level feature vectors were fused with the Child-Pugh classification to enable multimodal feature representation. In the present study, we employed a late fusion (or post-network fusion) approach based on vector concatenation (Supplementary material; Supplementary Figure 1). Finally, the fused features were passed through a regression head to generate the predictive HVPG value, which was subsequently mapped to a predefined HVPG interval for clinical interpretability. Besides, in order to mitigate data imbalance, the model integrated the Boundary-Aware Weighted loss into its optimization objective (Supplementary material). Optimization was performed using Adam with an initial learning rate of 0.001, training for 50 epochs with a batch size of 32.

Figure 1 Image pre-processing and model establishment. A: Region of interests delineation on computed tomography images. a: The first hepatic hilum section; b: The second hepatic hilum section; c: The section of the splenic vein merging into the portal vein; d: The section with the largest spleen area; B: Data rearrangement strategy was adopted by permuting and combining the computed tomography images of four anatomical sections in each patient; C: The pre-processed images were converted into vectors and subjected to the visual prompt tuning transformer network for feature extraction. Transfer learning indicated that the transformer network parameters were frozen and that only the prompt tokens were trainable; D: The extracted feature vectors underwent multimodal feature fusion, regression analysis, and finally output of the predictive hepatic venous pressure gradient values and classification interval. PV: Portal vein; IVC: Inferior vena cava; HV: Hepatic vein; SV: Splenic vein; PRT: Prompt token; CLT: Class token; VPT: Visual prompt tuning; HVPG: Hepatic venous pressure gradient.

Robustness of the HVPG prediction model

Model training was conducted using repeated five-fold cross-validation. The original dataset was randomly divided once into five equal folds via computer-generated random numbers without repeated, stratified sampling or optimization procedure. Five-fold cross-validation was subsequently performed by sequentially using one fold as the internal testing set and the remaining four folds as the training set, thus generating a total of five models for non-invasive prediction of HVPG. The overall prediction accuracy of them were compared to investigate the robustness of predicting the HVPG.

Comparison with other conventional non-invasive methods

Many non-invasive HVPG prediction methods have been reported, including imaging-based parameters such as LSM[13], serum-based indices such as aspartate aminotransferase-to-alanine aminotransferase ratio[14], aspartate aminotransferase-to-platelet count ratio index[15], CSPH risk score[16], fibrosis index based on four factors[17], fibrosis index[18], gamma-glutamyl transferase-to-platelet count ratio[19], King’s score[20] and Lok score[21], and novel composite models such as ANTICIPATE model[22] (Supplementary material; Supplementary Table 1). Our model was also compared with the aforementioned conventional non-invasive methods with respect to the prediction of different HVPG thresholds.

Statistical analysis

Continuous variables were expressed as mean ± SD or median (interquartile range) and were analyzed via analysis of variance (ANOVA) or the Kruskal-Wallis test. Categorical variables were presented as n (%) or frequencies, which were compared via χ² test or Fisher’s exact test. Post-hoc pairwise comparisons with multiple comparison adjustment were subsequently performed to identify the specific groups contributing to the overall significant differences. Pearson’s correlation coefficient analysis was used to assess the correlation between the model predictive values and the actual HVPG values. Bland-Altman analysis, error-based metrics such as mean absolute error and root mean squared error, and concordance correlation coefficient were calculated to further assess the agreement between the model predictive and the actual HVPG values. The diagnostic performance was evaluated by sensitivity, specificity, positive predictive value, negative predictive value and area under the receiver operating characteristic curve (AUC). Statistical analysis was performed by using GraphPad Prism Version 10.3.1 and SPSS Version 27.0.1.0. A two-tailed P value of less than 0.05 was considered statistically significant.

RESULTS

Participants and characteristics

A total of 607 patients with chronic liver diseases who received the HVPG measurement from September 2017 to August 2023 were screened. After excluding 248 patients, 359 patients were ultimately enrolled in this study. Of them, 228, 57 and 74 patients were assigned to the training set, the internal testing set, and the temporally independent validation set, respectively (Figure 2). The baseline characteristics of the three datasets were listed in Table 1. Hepatitis B was the most common underlying liver disease in all datasets, and the majority of patients had Child-Pugh classification of either A or B. In terms of the HVPG distribution, 46 of 359 (12.8%) HVPGs were less than 10 mmHg in all datasets, although the majority of HVPGs fell into the interval of ≥ 10 mmHg. EGVB remained the most common complication of portal hypertension, and 61.3% of patients received one or more sessions of endoscopic therapy. Besides, the overall analysis showed significant differences in age, history of EGVB, platelet counts, albumin, and prothrombin time among the three datasets. Subsequent post-hoc pairwise analysis revealed that only platelet counts and history of EGVB presented significant differences in the training and internal testing sets, whereas all other baseline characteristics were comparable among the three datasets after randomization (Supplementary Figure 2).

Figure 2 Flowchart of the study. TIPS: Transjugular intrahepatic portosystemic shunt; SAE: Splenic artery embolization; CT: Computed tomography; HVPG: Hepatic venous pressure gradient.

Table 1 Baseline characteristics of patients in all datasets, mean ± SD/n (%).

	Training set (n = 228)	Internal testing set (n = 57)	Temporally independent validation set (n = 74)	P value
Age (years)	51.5 ± 12.8	55.9 ± 15.0	55.2 ± 9.7	0.016
Gender
Male	148 (64.9)	31 (54.4)	42 (56.8)	0.218
Female	80 (35.1)	26 (45.6)	32 (43.2)
Etiology
Hepatitis B	113 (49.6)	18 (31.6)	39 (52.7)	0.152
Alcoholic	24 (10.5)	8 (14.0)	11(14.9)
Hepatitis C	8 (3.5)	1 (1.8)	2 (2.7)
Autoimmune	42 (18.4)	15 (26.3)	15 (20.3)
Others	41 (18.0)	15 (26.3)	7 (9.4)
Child-Pugh class
Class A	96 (42.1)	22 (38.6)	23 (31.1)	0.325
Class B	101 (44.3)	26 (45.6)	43 (58.1)
Class C	31 (13.6)	9 (15.8)	8 (10.8)
HVPG (mmHg)	17.9 ± 7.3	16.6 ± 10.3	15.9 ± 6.1	0.064
HVPG ≤ 5 mmHg	13 (5.7)	3 (5.3)	5 (6.8)
5 mmHg < HVPG < 10 mmHg	15 (6.6)	4 (7.0)	6 (8.1)
10 mmHg ≤ HVPG < 12 mmHg	10 (4.4)	3 (5.3)	3 (4.1)
12 mmHg ≤ HVPG < 16 mmHg	33 (14.5)	8 (14.0)	10 (13.5)
16 mmHg ≤ HVPG < 20 mmHg	50 (21.9)	14 (24.6)	15 (20.2)
HVPG ≥ 20 mmHg	107 (46.9)	25 (43.8)	35 (47.3)
Ascites
Yes	112 (49.1)	28 (49.1)	45 (60.8)	0.201
No	116 (50.9)	29 (50.9)	29 (39.2)
EGVB
Yes	199 (87.3)	38 (66.7)	62 (83.8)	< 0.001
No	29 (12.7)	19 (33.3)	12 (16.2)
Endoscopy therapy
No	77 (33.8)	29 (50.9)	33 (44.6)	0.109
1-2 sessions	115 (50.4)	21 (36.8)	32 (43.2)
3 and more sessions	36 (15.8)	7 (12.3)	9 (12.2)
Liver stiffness (kPa)	17.8 ± 9.0	15.2 ± 7.5	18.5 ± 7.3	0.186
Laboratory tests
Platelet (109/L)	71 ± 45.3	88 ± 62.6	70 ± 34.8	0.030
Total bilirubin (μmol/L)	27.9 ± 30.6	36.9 ± 47.5	27.4 ± 25.8	0.157
ALT (U/L)	37 ± 47.3	29 ± 27.7	35 ± 41.4	0.402
AST (U/L)	49 ± 52.2	39 ± 37.3	49 ± 49.5	0.396
ALP (U/L), median (IQR)	84 (56.8)	84 (87.0)	89 (60.5)	0.604
γ-GGT (U/L), median (IQR)	39 (59.8)	31 (65.0)	46 (85.5)	0.440
Albumin (g/L)	35.1 ± 6.2	35.7 ± 7.4	32.4 ± 5.1	0.002
Globulin (g/L)	28.5 ± 7.7	28.0 ± 7.3	30.6 ± 8.2	0.087
PT (seconds)	14.7 ± 2.4	14.0 ± 2.5	15.3 ± 2.9	0.018
INR	1.3 ± 0.2	1.2 ± 0.2	1.3 ± 0.3	0.080
Blood ammonia (μmol/L)	69.4 ± 36.7	58.8 ± 32.2	65.7 ± 36.9	0.220

HVPG: Hepatic vein pressure gradient; EGVB: Esophagogastric variceal bleeding; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; ALP: Alkaline phosphatase; γ-GGT: Gamma-glutamyl transferase; PT: Prothrombin time; INR: International normalized ratio; IQR: Interquartile range.

Correlation and agreement between the predictive and actual HVPGs

Figure 3 illustrated the correlation and agreement between the predictive and the actual HVPG values in the training, internal testing, and temporally independent validation sets. Significant correlations were observed in all three datasets (P < 0.001), with Pearson’s correlation coefficients of 0.949 [95% confidence interval (CI): 0.948-0.951], 0.741 (95%CI: 0.707-0.771) and 0.717 (95%CI: 0.683-0.747), respectively. Bland-Altman analysis revealed low mean biases occurred between the predictive and the measured HVPG values, suggesting minimal systematic overestimation or underestimation of the prediction model, although the 95% limits of agreement became wider in the internal testing and temporally independent validation sets compared with the training set. Furthermore, the model achieved mean absolute errors of 1.75, 4.66 and 3.69 mmHg, alongside root mean squared errors of 2.65, 6.25 and 5.32 mmHg in the training, internal testing and temporally independent validation sets, respectively. The corresponding concordance correlation coefficients were 0.947 (95%CI: 0.945-0.949), 0.684 (95%CI: 0.650-0.715) and 0.702 (95%CI: 0.669-0.732), indicating a moderate to strong clinical agreement with limited large deviation errors.

Figure 3 Correlation and agreement between the predictive and actual hepatic venous pressure gradient values. Scatterplot and Bland-Altman analysis showed the correlation and agreement between the actual hepatic venous pressure gradient and the predictive hepatic venous pressure gradient values. A and B: Training set; C and D: Internal testing set; E and F: Temporally independent validation set. HVPG: Hepatic venous pressure gradient; LOA: Limits of agreement.

Prediction accuracies on HVPG classification intervals with different granularities

The overall prediction accuracies referred to the concordance rates, representing the proportion of cases where both the predictive HVPG values and the actual HVPG values fell into the exact same predefined HVPG classification intervals, which were listed in Table 2. In the training set, the average overall prediction accuracies of the five-fold model for binary, ternary and quinary classification intervals were 93.4%, 89.2% and 75.2%, respectively. In the internal testing set, the average overall prediction accuracies of the five-fold model for binary, ternary and quinary classification intervals were 91.5%, 83.9% and 65.1%, respectively. Furthermore, in the temporally independent validation set, the average overall prediction accuracies of the five-fold model for binary, ternary and quinary classification intervals were 90.8%, 80.4% and 64.0%, respectively. As the granularity of the HVPG classification intervals increased, the overall prediction accuracy of the model decreased.

Table 2 Overall prediction accuracy of the model at different hepatic venous pressure gradient classification intervals.

	Binary classification interval					Ternary classification interval					Quinary classification interval
	Set 1	Set 2	Set 3	Set 4	Set 5	Set 1	Set 2	Set 3	Set 4	Set 5	Set 1	Set 2	Set 3	Set 4	Set 5
Training set	91.8	93.1	91.9	94.6	95.4	88.1	87.3	85.3	91.4	94.0	73.1	71.9	68.8	77.4	85.0
Internal testing set	89.9	92.3	89.1	93.9	92.4	85.2	79.7	84.1	82.8	87.6	62.9	63.7	64.5	65.7	68.6
Temporally independent validation set	90.4	91.3	88.5	92.3	91.3	80.8	75.0	82.7	78.8	84.6	60.6	61.5	64.4	63.5	70.2

Diagnostic performance in different binary HVPG classification intervals

As a result, our model displayed high sensitivity and high specificity in predicting the HVPG with different binary intervals in the training set. In both the internal testing set and the temporally independent validation set, as the HVPG thresholds increased, the model maintained stable sensitivity, with gradually improved specificity (Table 3). Our model exhibited promising diagnostic performance in the prediction of portal hypertension (HVPG > 5 mmHg), CSPH (HVPG ≥ 10 mmHg), and EGVB (HVPG ≥ 12 mmHg). Besides, the diagnostic performances of the ANTICIPATE model across various HVPG thresholds were consistently superior to that of serum-based indices and mostly outperformed isolated LSM (Supplementary Table 2). However, the receiver operating characteristic curves of our transformer model still surpassed those of conventional non-invasive methods in terms of the prediction of the HVPG (Figure 4).

Figure 4 Receiver operating characteristic curves of our hepatic venous pressure gradient prediction model and other conventional non-invasive models. HVPG: Hepatic venous pressure gradient; CT: Computed tomography; LSM: Liver stiffness measurement; AAR: Aspartate aminotransferase-to-alanine aminotransferase ratio; APRI: Aspartate aminotransferase-to-platelet count ratio index; CSPH: Clinically significant portal hypertension; FIB-4: Fibrosis index based on four factors; GPR: Gamma-glutamyl transferase-to-platelet count ratio.

Table 3 Diagnostic performance of the model for different hepatic venous pressure gradient thresholds.

HVPG	Group	Sensitivity (95%CI)	Specificity (95%CI)	PPV (95%CI)	NPV (95%CI)	AUC (95%CI)
> 5 mmHg	Training set	0.999 (0.998, 1.000)	0.817 (0.798, 0.834)	0.941 (0.935, 0.947)	0.998 (0.994, 0.999)	0.999 (0.999, 1.000)
	Internal testing set	0.990 (0.968, 0.997)	0.389 (0.183, 0.639)	0.964 (0.935, 0.981)	0.700 (0.354, 0.919)	0.956 (0.922, 0.991)
	Temporally independent validation set	0.967 (0.901, 0.992)	0.500 (0.223, 0.777)	0.937 (0.862, 0.974)	0.667 (0.309, 0.910)	0.951 (0.911, 0.991)
≥ 10 mmHg	Training set	0.986 (0.982, 0.989)	0.970 (0.963, 0.976)	0.983 (0.979, 0.987)	0.974 (0.968, 0.980)	0.998 (0.998, 0.999)
	Internal testing set	0.992 (0.970, 0.999)	0.519 (0.380, 0.655)	0.909 (0.868, 0.938)	0.933 (0.765, 0.988)	0.934 (0.905, 0.962)
	Temporally independent validation set	0.940 (0.859, 0.978)	0.857 (0.626, 0.962)	0.963 (0.888, 0.990)	0.783 (0.558, 0.917)	0.962 (0.927, 0.996)
≥ 12 mmHg	Training set	0.986 (0.982, 0.989)	0.990 (0.985, 0.993)	0.993 (0.991, 0.995)	0.978 (0.972, 0.983)	0.999 (0.999, 1.000)
	Internal testing set	0.996 (0.972, 1.000)	0.640 (0.531, 0.737)	0.875 (0.827, 0.912)	0.983 (0.895, 0.999)	0.978 (0.965, 0.991)
	Temporally independent validation set	0.912 (0.823, 0.961)	0.958 (0.769, 0.998)	0.986 (0.917, 0.999)	0.767 (0.573, 0.894)	0.972 (0.943, 1.000)
≥ 16 mmHg	Training set	0.938 (0.929, 0.946)	0.967 (0.961, 0.972)	0.958 (0.951, 0.965)	0.951 (0.944, 0.957)	0.993 (0.992, 0.994)
	Internal testing set	0.941 (0.899, 0.967)	0.853 (0.762, 0.914)	0.937 (0.894, 0.964)	0.862 (0.772, 0.921)	0.955 (0.922, 0.988)
	Temporally independent validation set	0.905 (0.809, 0.958)	0.833 (0.645, 0.937)	0.931 (0.839, 0.974)	0.781 (0.596, 0.901)	0.932 (0.877, 0.987)
≥ 20 mmHg	Training set	0.906 (0.891, 0.919)	0.969 (0.964, 0.973)	0.899 (0.883, 0.912)	0.971 (0.967, 0.975)	0.992 (0.991, 0.993)
	Internal testing set	0.793 (0.730, 0.845)	0.875 (0.796, 0.927)	0.920 (0.867, 0.954)	0.700 (0.616, 0.773)	0.895 (0.849, 0.940)
	Temporally independent validation set	0.750 (0.575, 0.873)	0.912 (0.811, 0.964)	0.818 (0.639, 0.924)	0.873 (0.768, 0.937)	0.904 (0.846, 0.961)

HVPG: Hepatic venous pressure gradient; PPV: Positive predictive value; NPV: Negative predictive value; AUC: Area under the receiver operating characteristic curve; CI: Confidence interval.

DISCUSSION

In our study, by combining the transformer network with transfer learning, we successfully developed a non-invasive model based on CT imaging features to classify and predict the HVPG. Previous studies constructed non-invasive HVPG prediction models from radiological images via machine learning, such as CNN-based algorithms. For example, Liu et al’s study[8] including 679 participants developed a deep CNN model for identifying CSPH with an AUC value of 0.90. Yamamoto et al’s[23] deep learning model based on CNN architecture achieved a sensitivity of 0.867 and specificity of 0.615 for CSPH, respectively. However, the CNN models were usually designed for classification tasks, which converted continuous values into discrete categories according to binary outcomes. While this strategy is effective for identifying clinically relevant cut-offs, it inevitably results in the loss of quantitative information. In contrast, our transformer-based model was originally developed as a regression framework to directly predict continuous HVPG values. The diagnostic performances across different HVPG thresholds were subsequently evaluated in the post-hoc analyses based on the predictive HVPG values as well, which preserved the full spectrum of portal venous pressure information and enabled both quantitative HVPG estimation and clinically relevant risk stratification. On the other hand, unlike CNNs that primarily extract local image features through convolutional operations, the transformer network adopts a self-attention mechanism to be capable of directly computing relationships between any two positions in an image, which allows the model to not only capture local details but also efficiently process global contextual information. In addition, the transformer network exhibits notable flexibility in handling heterogeneous data structures, making it well suited for integrating multimodal information[11]. Taking advantage of the transformer network, our model has a good correlation between the predictive HVPG and the actual HVPG values, with Pearson’ correlation coefficients of 0.741 in the internal testing set and 0.717 in the temporally independent validation set, and showed a moderate clinical agreement with limited large deviation errors.

However, the transformer network typically contains a large number of parameters and requires substantial computational resources and training time. Given the relatively limited CT image dataset available for training in this study, we adopted the transfer learning strategy, first pretraining on a large dataset and then fine-tuning on another small dataset, which effectively decreased the risk of overfitting, enhanced the model generalizability, reduced the computational load and storage requirements, shortened the training time, and finally improved the overall training efficiency[24]. The advantages of transfer learning ensured that the small training dataset in this study (with a sample size of 228) could produce a promising model to predict the HVPG.

The delineation of ROIs is a critical but challenging task for the establishment of a machine learning model based on radiological images because there are many images in each patient, and ROIs delineation is both time- and energy-consuming. For instance, Liu et al’s study[9] revealed that the average time for a virtual HVPG calculation was about 2.5 hours, with 1.5 hours for the human processing and 1.0 hour for machine computation. To overcome this plight, we chose four pivotal sections to simplify the ROIs delineation. Considering that liver morphology significantly shifts under portal hypertension[25], the four sections that we chose for ROIs delineation comprised the first hilum, the second hilum, the merged site of the portal vein and the splenic vein, and the largest section of the spleen. These structures are easy to recognize and contain important information of the portal system status. Our results suggested that the transformer model built on four images per capital shared almost equal efficacy with those models built on hundreds of images per capital[7,9,26], which highlighted our model, as it greatly decreased the workload of ROIs delineation and made real-time use of the model possible on the clinical frontline.

In addition to classic CNN models, various radiomics machine learning models have also been reported in recent years. For instance, Liu et al’s[8] radiomic signature, termed rHVPG, which was based on the CT images of 385 patients, showed good performance in the detection of CSPH, with a C-index of 0.849. Another auto-machine learning CT radiomics model, called aHVPG, developed by Yu et al[27] explored the performance for HVPG quantitation and multistage assessment with HVPG ≥ 10 mmHg and above. However, these models still aimed to identify CSPH or higher HVPG thresholds, patients with normal portal venous pressure and mild portal hypertension acquired relatively limited attention. Distinct from these studies, we focused not only on CSPH but also on the whole spectrum of portal venous pressure. This study included patients whose HVPG varied from 1 to 37 mmHg, and specifically, 5.3%-6.8% of the HVPGs in each dataset fell within the interval of ≤ 5 mmHg while 12.3%-14.9% within 10 mmHg. The covering of the whole spectrum of portal venous pressure better reflected every stage of chronic liver diseases, from chronic hepatitis and mild/moderate fibrosis to compensated and decompensated cirrhosis. In terms of results, the prediction model can distinguish portal hypertension with a high sensitivity of about 96% but a relatively low specificity of about 50%, which may be due to the fact that even though we actively included patients with compensated cirrhosis, the proportion of this group of patients was still at disadvantage compared to the large number of cases with CSPH. However, it was an innovative attempt to evaluate the radiomics machine learning models in the discrimination of patients with and without portal hypertension based on the HVPG threshold of 5 mmHg. While for the prediction of CSPH carried out by most of previous studies[7,8,26], our model was able to achieve a sensitivity of 94% and a specificity of 86%, and its diagnostic performance was not weaker or even better than that of them, with an AUC value of 0.962 in the temporally independent validation set. More importantly, with regard to the HVPG values that have indicative effects on decompensated events such as bleeding (HVPG ≥ 12 mmHg) and death (HVPG ≥ 16, 20 mmHg)[28], our model had the capability of achieving excellent prediction and classification, which could assist clinicians in initially evaluating the risk of complications or identifying high-risk patients who require timely intervention. It is noteworthy that several baseline characteristics showed significant differences among the three datasets, while subsequent post-hoc pairwise analysis revealed that most of these differences were primarily attributable to comparisons involving the temporally independent validation set. Although statistically significant differences in platelet counts and history of EGVB were still observed in the training and internal testing sets after randomization, likely due to patients’ heterogeneity inherent to retrospective studies, our model mainly relied on extracting high-dimensional, invariant anatomical and radiomic features from CT images, and these clinical variables were not directly incorporated into the model establishment, suggesting a limited impact on the model performance.

Among the available non-invasive prediction methods for the HVPG, LSM by ultrasound elastography is recommended by Baveno VII[2] as an effective tool for identifying CSPH. Consistent with previous studies[5,29], LSM in our study had an average diagnostic performance, with an AUC of 0.632-0.897. Some studies suggest that LSM is easily affected by the etiology of cirrhosis and various extrahepatic factors, such as obesity, ascites, or collateral circulation in the decompensated stage, hindering stable predictive accuracy[30]. In contrast, our model trained from CT images did not seem to be influenced by the above factors and had better predictive accuracy than LSM. Furthermore, the serum-based indices could not show enough predictive accuracy, with AUCs ranging from 0.516-0.765. This is because, unlike our model, serum-based indices have little knowledge of organ and vascular structure changes during the evolution of portal hypertension. The majority of the serum-based indices pivot temporary variables, whereas CT images display long-term changes in the portal system, which would better carter to study the duration of chronic liver diseases. Beyond single-modality prediction models, novel composite prediction tools, such as the ANTICIPATE model, which integrates LSM with platelet counts to achieve a robust discrimination for CSPH with an AUC of 0.82-0.85, were reported recently[22]. Our study similarly demonstrated the ANTICIPATE model outperformed isolated LSM or serum-based methods in the evaluation of HVPG, but its prediction efficacy remained inferior to the transformer model.

However, there were still some limitations in this study. Firstly, data imbalance of the HVPG still presented in our study. Patients with the HVPG < 5 mmHg accounted for less than 10% of the datasets, which was not sufficient to predict normal or mild portal hypertension. Although we incorporated the Boundary-Aware Weighted loss function during the model training process to partially handling this issue, more patients with HVPG < 5 mmHg should be included to improve our model’s prediction performance in this subgroup in the future. Secondly, CNN models were not trained and evaluated on the same datasets, precluding a direct comparison between the transformer and CNN-based architectures. Future studies should explore comparisons with CNN-based models trained under identical conditions, as well as with emerging non-invasive and multimodal composite prediction models. Besides, another potential limitation is that CT images were converted to Joint Photographic Experts Group format and RGB encoding, which inevitably resulted in the loss of original Hounsfield Units information. Although this pre-processing operation was required to accommodate the pretrained transformer architecture and preserved the diagnostically relevant imaging features through standardized abdominal window settings, future models based on native digital imaging and communications in medicine data may further improve the prediction performance and biological interpretability. Finally, this was a single-center study, and the generalizability of our model needs to be externally validated in large-scale, multicenter medical facilities.

CONCLUSION

In conclusion, we successfully developed and validated a non-invasive regression machine learning model based on the transformer network to predict the HVPG values using contrast-enhanced CT images. Our model accurately evaluated the entire spectrum of portal venous pressure and outperformed conventional non-invasive tools. Furthermore, simplifying ROIs delineation to four pivotal sections significantly reduced image pre-processing workload, enabling real-time clinical application. In the future, multicenter studies including more patients with normal or mild portal hypertension are needed to address data imbalances and validate the model generalizability.

ACKNOWLEDGEMENTS

We were grateful to Professor Da-Qing Guo and his team at the School of Life Science and Technology, University of Electronic Science and Technology for their valuable support and collaboration in this work.