Abstract
Background
Non-invasive imaging methods are still lacking for evaluating bone changes in chronic kidney diseases (CKD).
Purpose
To investigate the feasibility of chest CT radiomics in evaluating bone changes caused by CKD.
Material and Methods
In total, 75 patients with stage 1 CKD (CKD1) and 75 with stage 5 CKD (CKD5) were assessed using the chest CT radiomics method. Radiomics features of bone were obtained using 3D Slicer software and were then compared between CKD1 and CKD5 cases. The methods of maximum correlation minimum redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO) were used to establish a prediction model to determine CKD. The receiver operating characteristic (ROC) curve was used to determine the performance of the model.
Results
Cases of CKD1 and CKD5 differed in 40 radiomics features (P <0.05). Using the mRMR and LASSO methods, five features were finally selected to establish a predication model. The area under the receiver operating characteristic curve of the model in the determination of CKD1 and CKD5 was 0.903 and 0.854, respectively, for the training and validation cohorts.
Conclusion
Chest CT radiomics is feasible in evaluating bone changes caused by CKD.
Keywords
Introduction
Chronic kidney disease (CKD) is a common disease affecting millions of people in China (1). Abnormal levels of calcium/phosphorus and secondary hyperparathyroidism result in bone metabolism disorder in CKD (2,3). Renal osteodystrophy (ROD) is frequently seen at the end stage of renal diseases (4). The decrease in density and strength of bone increased the risk of fracture for patients with CKD (5,6). Local bone destruction (called fibrous osteoarthritis) can be observed in some cases of CKD using computed tomography (CT) or X-ray, while osteoporosis is more frequently seen (7). CT attenuation decrease of bone can be seen in most patients with CKD5 (8).
It is well known that the CT attenuation of bone is influenced by bone trabecular density and yellow marrow content. The content of yellow marrow increases with age, which makes bone CT attenuation decrease. Elderly individuals usually have low bone CT attenuation because of decreased trabecular density and increased yellow marrow. CT density is an imperfect parameter for identifying cases of CKD, because low CT density can be seen in both cases of CKD and elderly individuals (9). Other CT parameters are urgently required for the evaluation of CKD.
In recent years, CT radiomics has been used for the evaluation of bone diseases (10). Hundreds of radiomics features can be extracted from the CT image of bone, and some features are useful to identify the disease. The least absolute shrinkage and selection operator (LASSO) regression method is frequently used to select features and establish a model (11). There are many publications reporting the value of LASSO in CT radiomics of bone diseases (12). However, to the best of our knowledge, no study uses CT radiomics to evaluate the change in bone microstructure caused by CKD. The aim of the present study was to investigate the feasibility of chest CT radiomics in the determination of stage 1 CKD (CKD1) and stage 5 CKD (CKD5).
Material and Methods
Participants
This is a retrospective study approved by the Institutional Review Broad of the hospital. Informed consent was obtained from patients before the study. The inclusion criteria were as follows: (i) patients with non–contrast-enhanced chest CT; and (ii) glomerular filtration rate (GFR) <15 mL/min or GFR ≥90 mL/min. In the study, each pair of CKD1 and CKD5 cases had the same age and sex. Unpaired patients with CKD were excluded from the study.
Images analysis
The source data of chest CT were transferred to a workstation with 3D-slicer (version 4.21). The radiomics features of bone were extracted by a radiologist with 7 years of experience in diagnosing bone diseases. The radiologist was blinded to the patients’ details. A circular region of interest (ROI) was drawn on the transverse images of the L1 vertebra. Radiomics features of bone were extracted with the function radiomics of 3D-slicer. In the study, the first order, glcm, gldm, glrlm, glszm, and ngtdm were selected. In total, 93 features were obtained for each bone in the study.
Statistical analysis
All statistical analyses were performed using Pycharm (Python 3.9). We assessed the differences between cases of CKD1 and CKD5 using the two-sample t-test for normal distribution data or the Mann–Whitney U test for non-normal continuous variables. Features with P values <0.05 were considered significant, which were further selected by the method of maximum correlation minimum redundancy (mRMR). The method of LASSO regression was then used for the features selected by mRMR. The training and validation cohorts were split into a ratio of 8:2. The receiver operating characteristic (ROC) curve was constructed for evaluating the performance of the model in the determination of CKD1 and CKD5. The area under the ROC curve (AUC) was calculated for the training and validation cohorts, respectively.
Results
A total of 150 individuals (72 men, 78 women; age range = 31–73 years; median age = 52 years) were included in the study. Between September 2017 and May 2022, 75 cases of CKD1 and 75 cases of CKD5 with non–contrast-enhanced chest CT were analyzed in the study. In total, 93 radiomics features of CT imaging of bone were successfully extracted.
Cases of CKD1 and CKD5 differed in 40 features (P <0.05). Using the mRMR method, 12 features were selected (Table 1). Using LASSO method (Figs. 1 and 2), five features were finally selected: DependenceNonUniformity; Contrast; Correlation; DependenceVariance; and DependenceEntropy.
The coefficients of the features varied with lambda. The number of selected features reduced with the increase of lambda. Finally, the coefficients of all features became zero.
The fivefold cross-validation of the least absolute shrinkage and selection operator algorithm. A vertical line was drawn at the optimal value.
The features with the top mRMR score.
mRMR, maximum correlation minimum redundancy.
There were 120 cases in the training cohort and 30 cases in the validation cohort. The AUC of the LASSO model in the determination of CKD5 and CKD1 was 0.903 in the training cohort. The AUC of the model was 0.854 in the validation cohort.
Discussion
The present study investigated the feasibility of chest CT radiomics in evaluating bone changes caused by CKD. The most important findings were as follows: (i) CKD1 and CKD5 bones differed in 40 radiomics features; and (ii) the LASSO model was useful in the determination of CKD1 and CKD5.
In cases of CKD, elevated parathyroid hormone releases more Ca2+ from bone to blood (13). The microstructure of CKD bone is thus different from that of normal bone (14). As mineral and bone disorders are the most serious in cases of CKD5, bone density and strength are expected to be lowest. Mineral and bone disorders are generally mild in cases of CKD1, so the bone density is nearly normal.
In recent years, CT radiomics provided a non-invasive way to study the microstructure of tissue (15). We thus tried to study the microstructure of CKD bone using CT radiomics, because some radiomics features are actually closely related to microstructure. In the present study, we compared as many as 93 parameters between CKD1 and CKD5 cases. Some features were found to be significantly different, indicating the difference in mineral disorder and microstructure.
Chest CT is one of the most widely used examinations in clinical practice (16). Patients with CKD sometimes need to undergo a chest CT scan to rule out possible lung infections (17). We thus collected 150 cases of CKD with non–contrast-enhanced chest CT, which involved the sternum, scapula, rib, and spine. Vertebrae seemed to be the best target in the study, because they are more important than other bones. Although lots of vertebrae can be seen on the chest CT, the L1 vertebra was investigated in the study because the ROI placement was relatively easy.
In regression models, we generally call the regression model with L1 regularization term LASSO regression (18). The value of LASSO in selecting features has been proved by many publications (19). In the present study, the AUC of the LASSO model was 0.903 for the training cohort and 0.854 for the validation cohort, indicating effective avoidance of overfitting. Thus, the prediction accuracy was significantly improved by using the LASSO method.
Bone CT density varies with age for a person. Old people generally have lower CT density than young people, and they may differ in bone microstructure. The radiomics features of CT imaging of bone may vary with age. Male and female patients of the same age may differ in bone microstructure, as well as some radiomics features. In the present study, cases of CKD1 and CKD5 were strictly paired in age and sex, so the difference in radiomics was only caused by CKD stage.
The present study has some limitations. First, invasive bone biopsy pathology was not used as the reference standard in detecting microstructure change in CKD. As the L1 vertebra was located deep in the body, a bone biopsy was not ethically permitted. Second, this is a single-center study with a moderate sample size. Large multicenter studies are required to validate our results. However, we at least introduce a new way of differentiating CKD5 and CKD1. Third, only the L1 vertebra was studied using radiomics. The other vertebrae should be investigated in future studies.
In conclusion, chest CT radiomics is feasible in evaluating the bone changes caused by CKD.
Footnotes
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.