Improved identification of tumors in 18F-FDG-PET examination by normalizing the standard uptake in the liver based on blood test data

Medical imaging is subjectively or quantitatively evaluated in clinical practice [1]. Positron emission tomography (PET)/computed tomography (CT) is a widely used minimally invasive imaging modality for evaluating various neoplasms and other diseases of the human body [2]. The radioactive tracer ¹⁸F fluro-2- deoxy-D-glucose (¹⁸F-FDG) is used in combination with PET for diagnostic purposes to identify tissues with altered glucose metabolism [3]. Standardized uptake value (SUV) is a semi-quantitative measurement of ¹⁸F-FDG tracer uptake in tissues [4]. The SUV is influenced by several variables, including weight, sex, body mass index (BMI), plasma glucose level, duration of uptake phase, partial-volume effects, and recovery coefficient [1, 4].

The liver is a useful organ for evaluating FDG uptake in quality control [5], therapy assessment [6], and prognosis [7]. Owing to its large size, it is clearly seen on FDG-PET images. The liver may exhibit increased FDG uptake even in the absence of malignancy. Various background SUVs can contribute to the failure to detect liver tumors or cancer. In addition, a genuine hepatic lesion with moderately elevated uptake may be unnoticed. Several clinical variables have been reported to have an impact on hepatic FDG absorption [8]. Therefore, it is necessary to normalize them for easier detection of tumors by physicians, for example, by simply thresholding the “normalized” SUV at its maximum.

Clinical variables for the diagnosis and treatment of liver disease are often obtained from blood tests, commonly referred to as liver function tests. Abnormal blood test results are often the first sign of liver diseases, such as cirrhosis, and are sometimes used as a guide for PET/CT imaging of the abdomen. Serum indicators, such as aspartate aminotransferase (AST), alanine aminotransferase (ALT), alkaline phosphatase (ALP), bilirubin, and albumin, are commonly used to screen for liver abnormalities [9]. Moreover, age [10], blood glucose level [11], and BMI [12] have all been shown to affect liver FDG absorption. Abdominal CT is the first imaging modality used when imaging is necessary to examine anomalies [9]. We searched PubMed using the keywords “Diagnosis of FDG/PET,” “blood test results,” and “Z-score.” No previously published articles related to the objectives of the present study were identified.

PET/CT-derived SUVs of healthy background liver tissues are often used as a reference to characterize anomalies and assess a tumor’s response to treatment. Therefore, we aimed to enhance the diagnostic ability of FDG-PET by normalizing the liver SUV in each patient. We estimated the average SUV (and standard deviation) of the liver from various non-image-derived variables using machine learning-based methods. We also established a fully automatic method to create a personally normalized Z-score map of the liver SUV. The Z-score method is known as the standard score or standardized value method. It is a statistical technique used to assess how far a particular data point is from the mean of a dataset, and it is measured in terms of statistical deviations. The Z-score helps to understand how extreme or unusual a data point is within a distribution. Moreover, we aimed to establish a least absolute shrinkage and selection operator (LASSO)-based estimation formula for normalizing individual hepatic SUV maps. This regression method is primarily used for predictive modeling and regression analysis. The LASSO regression is commonly used in situations where there are many features but not all are important for the prediction. Moreover, this model helps balance model complexity and predictive accuracy by encouraging sparse models and automatically selecting features in the presence of high-dimensional data.

Using the proposed method, the SUV map can be replaced with a normalized Z-score map for daily image interpretation by physicians. Hopefully, this will help detect and diagnose hepatic masses more accurately and in a more standardized manner. We also analyzed the receiver operating characteristic (ROC) curve results of subjects with and without hepatic tumors and compared them with the ROC curve of normal SUV.

A total of 7744 patients were selected for further analysis from 7773 first time visited patients between the time duration of November 2006 and November 2017. We present the selection procedure in Fig. 1. After selection of the sample, we randomly divided the dataset into training set (n = 3872) and test set (n = 3872). We calculated the mean and standard deviation as a basic statistical result of demographic characteristics and blood test data of the training dataset, test dataset, and total cases (n = 7744). For more details, see Table 1. A true difference was found in mean result between the training and test datasets. More specifically, ALT and AST show a significant true difference in mean between the training and test datasets.

We also investigated the focal lesions with local SUV elevation in the abnormal cases. We present the pathologies of all 13 abnormal cases in Table 2. These pathological diagnoses were confirmed by a hospital physician. Among the abnormal cases, hepatocellular carcinoma was found in 5 cases. In four cases, we found multiple number of tumors. Besides this tabular pathological information, we show a pictorial representation (Fig. 2) of the Z-score maps obtained using our anomaly-detection method. We present three cases as an example out of 13 abnormal cases. In this figure, the original images (PET/CT) are presented in the first and second columns. The Z-score maps and SUV maximum intensity projection figures are presented in the third and fourth columns. And in the last column, Z-score maximum intensity projection in liver was presented. This pictorial presentation of Z-score and SUV itself result magnifies the improved identification method of Z-score for abnormalities detection comparing the SUV result.

We found almost the same scale and same number of bins when we presented a histogram of SUV mean and estimated mean Z-score for normal cases; see Fig. 3a and c. For abnormal cases, the SUV mean and estimated mean Z-score gave a right-skewed result; see Fig. 3b and d. We also present a scatter plot of normal and abnormal cases in the same Figure. For normal cases in scatterplot, we can explain a null relationship between the SUV mean and estimated mean Z-score (Fig. 3e). By contrast, the abnormal cases explain a positive linear relationship between the two variables (Fig. 3f). Moreover, we presented the result of ROC curve of the maximum Z-score versus the SUV_max in Fig. 4. In Fig. 4, the AUROC value of the Z-score_max was higher (0.917) than that of the SUV_max (0.890). The Z-score map interpreted the anomaly detection approach as being higher (0.91) than the SUV_max (0.89). The p-value was 0.21 using Delong’s test for the two correlated ROC curves.

Figure 5 presents the correlation plots of the variables used in this study. BMI and weight were positively correlated with the SUV mean. For example, the correlation value of BMI and SUV mean was 0.45, while that of weight and SUV mean was 0.4. Moreover, Triglyceride (0.32) and γGPT (0.24) showed a small positive correlation among the other biochemical markers.

Finally, Figure 6 shows the LASSO linear regression model with all the selected covariates fitted to the data. We selected the covariate lists and estimated the results using the test dataset. In the model, for the threshold lambda of 0.00046, the selected covariates with their slopes (in parenthesis) are: sex (0.05), weight (0.0033), BMI (0.02), blood glucose level (0.00006), triglyceride (0.0001), total cholesterol (0.0002), GTP (0.00002), total protein (0.037), creatinine (0.017), insulin (-3.202), albumin (0.025), and cholinesterase (0.00004) for SUV mean. The intercept of the model is 0.82. Moreover, the selected covariate with its slope (in parenthesis) for SUV standard deviation in the model was weight (0.0029) for a threshold lambda of 0.000079. The intercept of the model was 0.228. The left panel of Fig. 6 shows the fitted model for the SUV mean, while the right panel shows the fitted model for the SUV standard deviation. Similarly, the top and middle rows show the models for the training and test datasets of normal cases, respectively. The bottom row shows the fitted model for abnormal cases. For normal cases, we found a positive linear relationship in the plots. Based on two-fold cross-validation, we randomly divided the entire dataset into training and test datasets.

This study aimed to evaluate the importance of FDG-PET SUV in identifying the liver function associated with relevant blood enzymes. Using ¹⁸FDG, PET/CT plays a crucial role in the treatment of various malignancies, with applications that include but are not limited to diagnosis, staging, response assessment, restaging, and prognostication. The SUV_max of a tumor is a crucial semi-quantitative metric used to distinguish between benign and malignant tumors. However, another method called the Z-score is used to identify abnormalities in the human body. We used the Z-score map (Eq. 1) and presented the anomaly results in Fig. 2. In Fig. 4, the findings of the per-voxel ROC analysis indicate a superior AUROC for the Z-score compared to only the SUV_max value in the liver region. The main advantage of using the Z-score is that it normalizes the mean and standard deviation. This normalization is effective for the ROC. Our normalization techniques increased the clarity of abnormalities. Therefore, we believe that such normalization will be a key technique for improving the diagnostic effectiveness of measuring the maximum glucose consumption of tumors. In addition, we plan to analyze the correlation between the maximum Z-score and tumor pathologies (e.g., benign vs. malignant) in a future study.

Considering the importance of the FDG-PET SUV and after analyzing it in our study, we found a significant positive correlation between the mean SUV and BMI, body weight, triglyceride, cholinesterase, and total protein and a negative correlation between the mean SUV and amylase and HDL cholesterol. Body weight and BMI have been shown to affect the transport of FDG into tissues and blood [1]. According to this study, people with a higher BMI take up more normal blood and tissue FDG than those with a lower BMI. This is probably because fat does not store much FDG during fasting; therefore, more FDG is taken up by non-fatty tissues.

In addition to Pearson correlation, we used the LASSO regression, a well-known machine learning algorithm, to filter the variables. This finding is consistent with previous studies that found a positive effect of blood glucose levels on liver FDG uptake [1]. Moreover, this is the largest study in which a large number of subjects were used to estimate the regression formula using blood test variables.

As seen in this study and previous studies, increasing blood glucose levels reduces tumor FDG absorption while boosting normal blood and tissue uptake. In addition, we presented the standard deviation of SUV uptake to create each personalized Z-score map. The covariate weight showed a positive relationship of the model with a lambda value of 7.90e-5. Our results showed the diagnostic value of the training set in the model; however, it was not significantly different from that of the testing set, which may be related to the sample size (Fig. 6).

This study has some limitations, as we did not investigate all possible variables that could affect liver FDG uptake, such as hepatic steatosis, diabetes status, abnormal lipid profile, and liver function abnormalities. Moreover, the small number of patients with abnormal findings is a drawback of this study. A few FDG-PET and blood test examinations were performed on the same day. Furthermore, we estimated only the mean and standard deviation of the entire liver, and not the voxel-wise estimation of the SUV. However, the liver is a homogenous organ; therefore, we considered the voxel-wise estimation of the SUV to be redundant. The method we introduced here is very objective and more suitable for comparing regions and backgrounds than subjective methods, such as the Deauville score.

In conclusion, the AUROC for the Z-score is a better predictor of abnormalities than the SUV_max value in the liver region. Covariates, such as sex, weight, BMI, blood glucose level, triglycerides, total cholesterol, GTP, total protein, creatinine, albumin, insulin, and cholinesterase are significant predictors of FDG uptake in the liver. Our future studies will include an evaluation of the predictive value of the Z-score for tumor pathology estimation.