DOI: https://doi.org/10.21203/rs.3.rs-274189/v1
We retrospectively assessed 214 patients with chronic liver disease or liver cirrhosis who underwent magnetic resonance imaging (MRI) enhanced with gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid (Gd-EOB-DTPA) from August 2016 to May 2020 to evaluate the relationship between biochemical results that reflect liver function and hepatic enhancement. With the information gained we employed a machine learning approach with the K-Nearest Neighbor (KNN) algorithm to develop a predictive model for determining insufficient hepatic enhancement during the hepatobiliary phase (HBP) in Gd-EOB-DTPA-enhanced MRI. Using both quantitative and qualitative assessments, the total bilirubin (TB), albumin (Alb), prothrombin time-international normalized ratio, platelet, Child-Pugh score (CPS), and Model for End-stage Liver Disease Sodium (MELD-Na) score were related to decreased hepatic enhancement. In a multivariate analysis, TB and Alb were associated with insufficient enhancement (p < 0.001). The predictive model showed that a combination of a variety of biochemical parameters had better performance (accuracy = 82.8%, area under the curve (AUC) = 0.861) in predicting insufficient enhancement than either the CPS (accuracy = 79.5%, AUC = 0.845) or the MELD-Na score (accuracy = 80.8%, AUC = 0.821). By using a machine-learning-based predictive model with the KNN algorithm, radiologists can predict insufficient hepatic enhancement during HBP in advance and adjust each patient's individually optimized MRI protocol.
The liver-specific contrast agent gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid (Gd-EOB-DTPA) has gained attention in recent years and has been included in a guideline on the management of hepatocellular carcinoma 1. This contrast medium allows for the assessment of tissue perfusion during hepatocyte-specific imaging. It helps in the detection and differentiation of focal liver lesions through a comparative evaluation between the hepatobiliary phase (HBP) and other vascular phases 2,3.
In patients with normal liver function, 50% of Gd-EOB-DTPA is uptaken by the hepatocyte through the organic anion-transporting polypeptides membrane transporter, and then secreted into the bile duct by an exporter transporter such as multidrug resistance-associated protein 2. The other half of the Gd-EOB-DTPA passes through the vessel after extracellular space distribution and is excreted through the kidney 4–6. As a result, in patients with impaired liver function, such as liver cirrhosis patients, the number of functioning hepatocytes decreases, resulting in a decrease in hepatocyte uptake of Gd-EOB-DTPA 7–9. Insufficient enhancement of liver parenchyma has been reported to reduce diagnostic accuracy 7,10. Therefore, it would be important to predict whether there will be insufficient enhancement before taking magnetic resonance imaging (MRI). There have been studies to predict insufficient enhancement during the HBP using the Child-Pugh score (CPS), the Model for End-stage Liver Disease (MELD) score, and a combination of other biochemical parameters 9,11. However, research conducted in conjunction with machine learning has not yet been reported to the best of our knowledge.
The purpose of this study is to evaluate the relationship between a variety of biochemical results reflecting liver function and hepatic enhancement and develop a predictive model for determining insufficient hepatic enhancement during HBP in Gd-EOB-DTPA-enhanced MRI using a machine learning approach with the K-Nearest Neighbor (KNN) algorithm.
All MRI examinations were performed with two clinical 3T MRI systems (Magnetom Skyra; Siemens Healthcare, Erlangen, Germany, and Magnetom Vida; Siemens Healthcare, Erlangen, Germany). A combination of body and spine coil elements was used for signal acquisition, with patients holding their breath in the supine position. The HBP images were obtained 20 min after administration of Gd-EOB-DTPA (Primovist, Bayer Healthcare, Berlin, Germany).
The parameters of the HBP sequence were: repetition time (TR), 4.2 ms; echo time (TE), 1.2 ms; flip angle (FA), 15’; matrix size, 256 ⋅ 187; field of view (FOV), 308 ⋅ 380 mm2; slice thickness, 3 mm; acquisition time, 15 s. Gadoxetic acid (0.025 mmoL/kg body weight) was administered via bolus injection (flow rate: 1 mL/s, flushed with 20 mL normal saline).
Analysis of MRI was retrospectively performed by three radiologists (WJY and BJ with > 10 years and KJS with 2 years of experience in interpretation of Gd-EOB-DTPA-enhanced MRI) who were kept unaware of related clinical information.
For quantitative analysis, regions of interest (ROI) analyses were performed. We drew 5–15 mm ROIs in diameter on the largest transversal slice of the liver of the HBP images, avoiding blood vessels, focal liver lesions, and artifacts. One ROI was placed in the middle of the right main branch of the portal vein and four ROIs were placed in the right lobe (anterior and posterior segments), and left lobe (medial and lateral segments). The liver-to-portal vein signal intensity ratio (LPR) was calculated by dividing the mean value of the signal intensity (SI) of the four liver parenchyma ROI by the SI of the portal vein. For qualitative assessment, visual scoring for relative hepatic enhancement relative to the portal vein on HBP images as liver-to-portal vein contrast (LPVC) was made using the following 5-level grading scale: 1, hyperintense; 2, slightly hyperintense; 3, isointense; 4, slightly hypointense; 5, hypointense 12. An LPVC score ≥ 3 was considered as an insufficient hepatic enhancement 11. Visual analysis was performed on different days at a 2-week interval from a quantitative assessment of SI for reducing recall bias, and the three reviewers were blinded to each other’s ROI measurements.
A variety of blood serum biochemical tests at the time of performing MRI, including alanine aminotransferase (ALT), aspartate aminotransferase (AST), gamma-glutamyl transpeptidase (GGT), alkaline phosphatase (ALP), albumin (Alb), total bilirubin (TB), platelet count (PLT), prothrombin time-international normalized ratio (PT-INR), activated partial thromboplastin time (aPTT), sodium (Na), and creatinine (Cr) were recorded. CPS and MELD-Na scores were also calculated for assessing hepatic function. The CPS was calculated from five variables including TB, Alb, PT, ascites status, and degree of encephalopathy 13. The equation for the MELD-Na score was MELD-Na = 11.2 × ln(INR) + 9.57 × ln(Cr, in milligrams per deciliter) + 3.78 × ln(TB, in milligrams per deciliter) + 6.43 – Na − [0.025 × (11.2 × ln(INR) + 9.57 × ln(Cr, in milligrams per deciliter) + 3.78 × ln(TB, in milligrams per deciliter) + 6.43) × (140 − Na)] + 140 14.
Continuous variables were expressed as the mean with standard deviation or the median with interquartile range. Categorical variables were expressed as number or frequency. Data were compared with the Student’s t-test, the nonparametric Mann–Whitney U test, or the Kruskal-Wallis test for continuous variables, and the Chi-squared test or the Fisher's exact test for categorical variables, as appropriate. Relationships between the visual assessment of hepatic enhancement and LPR and clinical factors were assessed using Pearson’s correlation coefficient (r) and Spearman’s rank correlation coefficient (ρ).
We performed uni- and multivariate logistic regression analyses to determine associates of insufficient hepatic enhancement by including the parameters that showed a significant difference in the univariate analyses. Biochemical parameters were used as variables, and calculated values such as CPS and MELD-Na were excluded.
These statistical analyses were performed using SPSS (version 21.0; SPSS, Inc., Chicago, IL, USA). A p-value < 0.05 was considered statistically significant. The predictive model for insufficient liver enhancement was developed using Matlab 2018b with the Statistical Machine and Deep Learning toolboxes. The utility of the machine learning algorithm for the prediction of insufficient hepatic enhancement was tested by omitting one cross-validation to calculate sensitivity, specificity, accuracy, and area under the curve (AUC) for CPS, MELD-Na, and the combination of biochemical parameters.
Patients and MRI measurements.
Table 1 summarizes the demographics, etiology of liver disease, and laboratory values of the 214 patients. These patients included 154 men (mean age, 61.45 ± 11.03 years) and 60 women (mean age, 67.32 ± 11.05 years). Hepatitis B virus was the most common cause of chronic liver disease (79.8%). The mean LPR was 1.84. Among 214 patients, 76 patients (35.5%) were classified as LPVC grade 1, 75 (35.0%) as LPVC grade 2, 57 (26.6%) as LPVC grade 3, six patients (2.8%) as LPVC grade 4, and there were no patients classified as grade 5.
characteristic | total (n = 214) | |
---|---|---|
Age, years* | 63.10 ± 11.32 | |
Sex† | male | 154 (71.9) |
female | 60 (28.0) | |
Etiology† | Hepatitis B | 126 (58.9) |
Hepatitis C | 28 (13.1) | |
Alcoholic liver disease | 49 (22.9) | |
primary biliary cirrhosis | 2 (0.9) | |
others‡ | 9 (4.2) | |
CPC† | A | 177 (82.7) |
B | 34 (15.9) | |
C | 3 (1.4) | |
CPS | 6 (5, 7) | |
MELD-Na | 8.5 (7, 12) | |
TB, ng/mL | 0.9 (0.6, 1.6) | |
Alb, ng/mL | 3.8 (3.28, 4.3) | |
PT-INR | 1.12 (1.05, 1.24) | |
aPTT, sec | 33.1 (29.9, 34.53) | |
AST, IU/L | 38 (29, 63.25) | |
ALT, IU/L | 31.5 (21, 50.25) | |
ALP, IU/L | 98 (78, 140.25) | |
GGT, IU/L | 60 (33.75, 126.25) | |
Cr, mg/dL | 0.8 (0.67, 0.96) | |
Na, mEq/L | 138 (136, 140) | |
PLT, ×1000/uL | 127.5 (80.75, 196.5) | |
LPVC† | 1 | 76 (35.5) |
2 | 75 (35) | |
3 | 57 (26.6) | |
4 | 6 (2.8) | |
5 | 0 (0) | |
Note. - Unless otherwise specified, data are median with interquartile range. | ||
* Data are presented as a mean ± standard deviation | ||
† Data are presented as number (%) of patients | ||
‡Including nonalcoholic steatohepatitis, autoimmune hepatitis, and cryptogenic liver disease. | ||
CPC = Child-Pugh class, CPS = Child-Pugh score, MELD-Na = Model for End-stage Liver Disease and sodium, TB = total bilirubin, Alb = albumin, PT-INR = prothrombin time-international normalized ratio, aPTT = activated partial thromboplastin time, AST = aspartate aminotransferase, ALT = alanine aminotransferase, ALP = alkaline phosphatase, GGT = gamma-glutamyl transpeptidase, Cr = creatinine, Na = sodium, PLT = platelet, LPR = liver-to-portal vein signal intensity ratio, LPVC = liver-to-portal vein contrast |
The ICC of measurement of SI among the three readers was 0.92 (95% confidence interval [CI]: 0.89–0.95), indicating a very strong positive correlation between the three readers’ measurements. There was substantial to almost perfect agreement (weighted κ = 0.7–0.96) in the pair-wise evaluation and substantial agreement (Fleiss κ = 0.736) among the three readers for classifying parenchymal enhancement grading. The visual assessment showed a significant correlation with quantitative hepatic enhancement with − 0.787 of ρ (p < 0.001; Fig. 1).
Correlation between MRI measurements and biochemical parameters.
Table 2 shows the relationships between LPR and the 5-level visual assessment and the biochemical parameters. Decreased serum levels of Alb (r = 0.518 for LPR, p < 0.001; ρ = -0.624 for the 5-level visual assessment, p < 0.001), PLT (r = 0.286, p < 0.001; ρ = -0.312, p < 0.001), and elevated TB (r = -0.381, p < 0.001; ρ = 0.460, p < 0.001), PT-INR (r = -0.335, p < 0.001; ρ = 0.442, p < 0.001), CPS (ρ = -0.482, p < 0.001; ρ = 0.643, p < 0.001), MELD-Na score (ρ = -0.427, p < 0.001; ρ = 0.535, p < 0.001) were significant factors related to decreased hepatic enhancement. Elevated AST (ρ = 0.311, p < 0.001), ALP (ρ = 0.166, p = 0.015), GGT (ρ = 0.241, p < 0.001) and decreased Cr (ρ = -0.262, p < 0.001) showed a significant relationship with decreased hepatic enhancement only for the 5-level visual assessment.
Liver-to-portal vein signal intensity ratio | Visual assessment for degree of hepatic enhancement | ||||||||
---|---|---|---|---|---|---|---|---|---|
Grade | |||||||||
correlation coefficient (r) | p-value† | 1 (n = 76) | 2 (n = 75) | 3 (n = 57) | 4 (n = 6) | p-value‡ | correlation coefficient (ρ) | p-value§ | |
TB | -0.381 | < 0.001 | 0.8 (0.6, 1) | 0.7 (0.5, 1.1) | 1.9 (1.15, 2.95) | 3.3 (1.78, 4.63) | < 0.001 | 0.460 | < 0.001 |
Alb | 0.518 | < 0.001 | 4.2 (4, 4.5) | 3.8 (3.4, 4.2) | 3.1 (2.75, 3.45) | 2.65 (2.45, 3.25) | < 0.001 | -0.624 | < 0.001 |
PT-INR | -0.335 | < 0.001 | 1.08 (1.02, 1.17) | 1.11 (1.03, 1.2) | 1.27 (1.15, 1.45) | 1.42 (1.13, 1.66) | < 0.001 | 0.442 | < 0.001 |
aPTT | -0.107 | 0.119 | 33.1 (30.43, 33.88) | 33.1 (29.2, 34.5) | 33.1 (30.5, 36.25) | 33.4 (30.28,35.05) | 0.397 | 0.089 | 0.197 |
AST | 0.027 | 0.693 | 32 (26.25, 46.75) | 37 (28, 65) | 51 (37, 83.5) | 60 (48, 211.75) | < 0.001 | 0.311 | < 0.001 |
ALT | 0.120 | 0.081 | 32 (22, 49.5) | 32 (20, 53) | 31 (20.5, 46) | 22.5 (13.5, 84.5) | 0.823 | -0.048 | 0.483 |
ALP | -0.072 | 0.298 | 95.5 (75.25, 128.75) | 92 (72, 131) | 110 (91.5, 156.5) | 115.5 (96.5, 169.75) | 0.026 | 0.166 | 0.015 |
GGT | -0.073 | 0.287 | 46.5 (30.25, 96.25) | 62 (33, 134) | 70 (44.5, 170.5) | 198.5 (120, 360) | 0.002 | 0.241 | < 0.001 |
Cr | 0.004 | 0.957 | 0.87 (0.75, 0.99) | 0.79 (0.68, 0.97) | 0.72 (0.6, 0.91) | 0.63 (0.46, 0.73) | 0.001 | -0.262 | < 0.001 |
Na | 0.028 | 0.679 | 138 (137, 140) | 139 (137, 141) | 137 (134.5, 139) | 140.5 (136.5, 141.25) | 0.008 | -0.122 | 0.076 |
PLT | 0.286 | < 0.001 | 164.5 (100.25, 221.75) | 146 (92, 215) | 91 (66, 129.5) | 82.5 (56.75, 117.25) | < 0.001 | -0.312 | < 0.001 |
CPS* | -0.482 | < 0.001 | 5 (5, 6) | 6 (5, 7) | 7 (7, 8.5) | 9 (8, 10.75) | < 0.001 | 0.643 | < 0.001 |
MELD-Na* | -0.427 | < 0.001 | 8 (7, 9) | 8 (7, 10) | 13 (10, 17) | 14.5 (11, 19.25) | < 0.001 | 0.535 | < 0.001 |
Note. - Unless otherwise specified, data are median with interquartile range. | |||||||||
* Data are presented as number (%) of patients. | |||||||||
† P value was calculated by Pearson’s correlation coefficient. | |||||||||
‡ P value was calculated by Kruskal-Wallis test. | |||||||||
§ P value was calculated by Spearman’s rank correlation coefficient. |
The univariate analysis showed that TB (odds ratio [OR] = 6.027, 95% CI: 3.44–10.55, p < 0.001), Alb (OR = 0.086, 95% CI: 0.04–0.17, p < 0.001), PT-INR (OR = 134.854, 95% CI: 21.64–840.33, p < 0.001), Na (OR = 0.91, 95% CI: 0.84–0.99, p = 0.02), and PLT (OR = 0.99, 95% CI: 0.99–0.99, p < 0.001) were significantly associated with insufficient hepatic enhancement during the HBP imaging (Table 3). The multivariate analysis revealed significant associations with TB (OR = 4.71, 95% CI: 2.2–10.01, p < 0.001) and Alb (OR = 0.12, 95% CI: 0.05–0.29, p < 0.001 [Table 3]).
Univariate analysis | Multivariate analysis | |||
---|---|---|---|---|
Odd ratio (95% confidence interval) | p-value | Odd ratio (95% confidence interval) | p-value | |
TB | 6.03 (3.44, 10.55) | < 0.001 | 4.71 (2.2, 10.01) | < 0.001 |
Alb | 0.086 (0.04, 0.17) | < 0.001 | 0.12 (0.05, 0.29) | < 0.001 |
PT-INR | 134.85 (21.64, 840.33) | < 0.001 | 0.2 (0.01, 5.06) | 0.325 |
aPTT | 1.02 (0.97, 1.07) | 0.465 | ||
AST | 1.00 (1, 1.01) | 0.100 | ||
ALT | 1 (1, 1) | 0.865 | ||
ALP | 1 (1, 1.01) | 0.199 | ||
GGT | 1 (1, 1) | 0.140 | ||
Cr | 1.01 (0.58, 1.76) | 0.982 | ||
Na | 0.91 (0.84, 0.99) | 0.020 | 1.04 (0.91, 1.19) | 0.551 |
PLT | 0.99 (0.99, 0.99) | < 0.001 | 1 (0.99, 1) | 0.529 |
Diagnostic performance of prediction of insufficient liver enhancement.
Diagnostic performance for the prediction of insufficient liver enhancement based on the KNN algorithm are presented in Table 4 and Fig. 2. The accuracies of the predictive model using KNN were 79.5% for CPS and 80.8% for MELD-Na. The accuracy of the predictive model from a combination of biochemical parameters had a higher accuracy of 82.8% than others (Fig. 3). The AUC of this combination of biochemical parameters showed the highest predictive ability with 0.861, followed by CPS (AUC = 0.845) and MELD-Na (AUC = 0.821). Regarding the 5-level grading of hepatic enhancement, a machine learning model with a KNN algorithm for classification as grade 1 achieved accuracies of 69.8% with CPS, 57.8% with MELD-Na, and 65.2% with a combination of biochemical parameters. For classification as grade 2, the accuracies were 58.6% with CPS, 60.1% with MELD-Na, and 57.7% with a combination of biochemical parameters. The classification accuracies for grade 3 and 4 were 78.0%, and 96.4% with CPS, 78.4%, and 97.0% with MELD-Na, and 80.8%, and 97.3% with a combination of biochemical parameters, respectively (Supplementary Information Table 1).
CPS | MELD-Na | Combination of multiple parameters | |
---|---|---|---|
Sensitivity, % | 95.4 ± 0.03 | 88.8 ± 0.04 | 95.6 ± 0.03 |
Specificity, % | 35.6 ± 0.15 | 62.4 ± 0.07 | 55.9 ± 0.09 |
Accuracy, % | 79.50 ± 0.06 | 80.8 ± 0.03 | 82.8 ± 0.04 |
AUC | 0.85 ± 0.05 | 0.82 ± 0.06 | 0.86 ± 0.05 |
Note. — Data are presented as a mean ± standard deviation. |
In this study we showed that laboratory tests reflecting hepatic function are closely related with hepatic enhancement during HBP. Using both quantitative and qualitative assessments, decreased serum levels of Alb, PLT, and elevated TB, PT-INR, CPS, MELD-Na score were related to decreased hepatic enhancement on HBP. Multivariate analyses revealed that increased TB and decreased Alb were significantly associated with decreased hepatic enhancement at HBP. We also used a machine learning algorithm to develop a predictive model for insufficient hepatic enhancement of HBP using a combination of biochemical parameters as well as CPS and MELD-Na, which are well-known scoring systems for reflecting reserved hepatic function. The prediction of insufficient hepatic enhancement during the HBP using KNN with a combination of biochemical parameters showed a higher accuracy and AUC than CPS or MELD-Na.
It would be useful if a simple visual MRI finding could determine the possibility of obscuring focal hepatic lesions due to insufficient enhancement of the background liver parenchyma. Measurements of LPR showed a strong relationship with LPVC grade, which was consistent results with a previous study by Tamada et al. 12. In this regard, we could easily identify insufficient hepatic enhancement during HBP. Furthermore, our results showed that both measurements of LPR and a 5-level assessment of hepatic enhancement provided similar patterns of association as the biochemical parameters.
Some previous studies have reported the association between liver function and insufficient enhancement during the HBP 9,11. Biochemical parameters commonly known to be associated with liver function are TB, ALT, ALP, GGT, Alb, PLT, and coagulation tests such as PT-INR 11,16−18. Our results were consistent with the previous studies. Insufficient hepatic enhancement during HBP was correlated with TB, Alb, PT-INR, platelet, and the 5-level degree of hepatic enhancement during HBP was correlated with TB, Alb, PT-INR, AST, ALP, GGT, Cr, and PLT levels. Multivariate analysis showed that Tb (OR = 4.71) and Alb (OR = 0.12) were the independent factors for predicting insufficient hepatic enhancement on HBP, which showed a stronger correlation than any other single parameter examined. Because GB-EOB-DTPA shares the same transport and excretion pathways as bilirubin, elevated serum TB is thought to reduce hepatocyte uptake and bile secretion of GB-EOB-DTPA 19. The decrease in serum Alb, which is known as a critical plasma protein produced by the liver, not only reflects the chronicity of liver disease but also has potential diagnostic value for determining a prognosis 20.
As shown in a previous study, many liver function tests and hepatic enhancement have a close relationship with HBP 21,22. However, because these liver function tests were all different, the combination of these parameters would be a better indicator than a single parameter 17. Several scoring systems have been accepted for assessment of reserved hepatic function in patients with liver cirrhosis, such as CPS, MELD, and MELD-Na scores 23–25. In our study, both CPS (r = -0.482, ρ = 0.643) and MELD-Na (r = -0.427, ρ = 0.535) were included among the high correlation factors concerning the LPR and the 5-level visual assessment of hepatic enhancement on HBP.
Machine learning makes it possible to train algorithms to discover and identify complex patterns and relationships within a variety of parameters by semi-automating the extraction of knowledge and insights from complex data. 26 In medicine, predictive studies based on machine learning are emerging and developed algorithms can be directly applied to patient care to improve the accuracy of predicting diseases and subsequent outcomes 27. Insufficient enhancement during the HBP can cause poor contrast differences between background liver parenchyma and a focal hepatic lesion without uptake of Gd-EOB-DTPA 7,28. If insufficient enhancement during the HBP can be predicted in advance, image quality can be improved by a delay of the time to obtain the HBP or modification of several sequence parameters such as flip angle and K-space 29,30. Therefore, it is clinically important to predict insufficient enhancement of liver parenchyma before performing Gd-EOB-DTPA-enhanced MRI. There have been studies to predict insufficient enhancement during the HBP using a traditional classification approach 9,11. Of our predictive models based on machine learning with the KNN algorithm, the model using a combination of biochemical parameters showed a higher accuracy of 82.8% and AUC of 0.86 than the models using CPS or MELD-Na. As mentioned above, because the various parameters reflecting liver function all work differently, a combination of the various parameters would more accurately determine hepatic enhancement on HBP. While the prediction of insufficient hepatic enhancement on HBP yielded a high diagnostic performance, classification of hepatic enhancement using a 5-level visual assessment showed incomplete results because of the small data set of unbalanced data.
There were some limitations in our study. First, this study was conducted retrospectively, and thus carries the potential for selection bias. Second, the model was developed from a small population in a single center and was not validated using another set of populations. Therefore, our results should be further validated through prospective studies in multiple institutions. Finally, we did not consider the degree of liver fibrosis, which also affects hepatic enhancement on HBP. Liver fibrosis was identified with an invasive biopsy procedure, which has the inherent problems of safety and some degree of sampling error. Nevertheless, research comparing the performance of the predictive models when integrating histopathologic results should be conducted in the future.
In conclusion, radiologists can predict insufficient hepatic enhancement during HBP in advance with a machine learning-based predictive model that uses the KNN algorithm to adjust each patient's individually optimized MRI protocol.
Acknowledgments
None
Conflict of interest statement:
The authors declare no conflicts of interest that pertain to this work.
Authors’ contributions:
*J.B and S.P contributed equally to this work as co-corresponding authors
J.B and S.P contributed to study concept and design. J.S.K, J.B and J.Y.W acquired, analyzed, interpreted the data. J.B and S.P performed statistical analysis and developed prediction mode. J.S.K and J.B drafted the manuscript. J.S.K, J.B, S.P and J.Y.W made critical revisions to the manuscript.