Access the full text.
Sign up today, get DeepDyve free for 14 days.
Purpose: Nephrologists have empirically predicted renal function from renal morphology. In diagnosing a case of renal dysfunction of unknown course, acute kidney injury and chronic kidney disease are diagnosed from blood tests and an imaging study including magnetic reso- nance imaging (MRI), and an examination/treatment policy is determined. A framework for the estimation of renal function from water images obtained using the Dixon method is proposed to provide information that helps clinicians reach a diagnosis by accurately estimating renal func- tion on the basis of renal MRI. Approach: The proposed framework consists of four steps. First, the kidney area is extracted by MRI using the Dixon method with a U-net by deep learning. Second, the extracted renal region is registered with the target mask. Third, the kidney features are calculated based on the target mask classification information created by a specialist. Fourth, the estimated glomerular filtration rate (eGFR) representing the renal function is estimated using a regression support vector machine from the calculated features. Results: For the accuracy evaluation, we conducted an experiment to estimate the eGFR when MRI was performed and the eGFR slope, which is the annual rate of decline in eGFR. When the accuracy was evaluated for 165 subjects, the eGFR was estimated to have a root mean square error (RMSE) of 11.99 and a correlation coefficient of 0.83. Moreover, the eGFR slope was estimated to have an RMSE of 4.8 and a correlation coefficient of 0.5. Conclusions: Therefore, the proposed method shows the possibility of estimating the prognosis of renal function based on water images obtained by the Dixon method. © The Authors. Published by SPIE under a Creative Commons Attribution 4.0 International License. Distribution or reproduction of this work in whole or in part requires full attribution of the original publication, including its DOI. [DOI: 10.1117/1.JMI.9.2.024501] Keywords: quantitative estimated glomerular filtration rate; magnetic resonance imaging; kidney. Paper 21162R received Jul. 1, 2021; accepted for publication Feb. 23, 2022; published online Mar. 15, 2022. 1 Introduction Chronic kidney disease (CKD) is defined as a persistent decrease in estimated glomerular fil- tration rate (eGFR) to <60 mL∕ min per 1.73 m , the presence of abnormalities suggestive of 1,2 renal injuries, such as proteinuria, or both. When CKD progresses, renal replacement therapy is required. Moreover, CKD is also important as an underlying condition related to arterioscle- rosis and immunodeficiency, so CKD can be seen to be related to a number of major causes of 1,2 death. CKD is a global medical problem affecting from 8% to 16% of the population worldwide. Due to the wide variety of causes of CKD, there is no specific therapeutic inter- vention, and it is necessary to detect the disease early and control the risk factors for kidney damage. *Address all correspondence to Masahiro Ishikawa, email@example.com Journal of Medical Imaging 024501-1 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging When a nephrologist treats a patient with renal dysfunction with an unknown clinical course, the nephrologist often refers to imaging findings for the kidney in addition to laboratory data and medical history. To this end, renal ultrasonography, abdominopelvic x-ray computed tomography, and renal magnetic resonance imaging (MRI) are useful. The nephrologist predicts the potential renal function suggested by the morphology and reflects this potential in the treatment planning. MRI has a particularly good resolution in terms of soft-tissue contrast and, with the use of appropriate imaging methods, it is possible to obtain detailed information on the internal struc- ture of the kidney, such as the corticomedullary border. Moreover, MRI can also provide data on the physiological aspects of the kidney. The T2* value of the blood oxygenation level-dependent (BOLD) method is an index of ischemia/hypoxia, which may lead to the progression of CKD, 5,6 and is significantly correlated with the rate of deterioration of CKD. The apparent diffusion coefficient value of the diffusion-weighted image is an index related to renal fibrosis and is significantly correlated with the pathological findings of renal biopsy. As described above, there are high expectations regarding MRI as a noninvasive and multifaceted kidney evaluation method, but MRI has one drawback in that there is no method available for the comprehensive quantification of images. Traditionally, medical images have been measured by the region of interest (ROI) method. In the ROI method, a rectangular or circular area is selected, and the average value of the signal intensities in the area is used as a representative value. Problems with this method include the possibility that arbitrary decisions made by the observer may involve difficulty in including location information and the fact that only a part of the image can be measured. Pruijm et al. proposed the 12-layer concentric objects (TLCO) method for analyzing the renal region by dividing this region into 12 layers. The TLCO method is a method of designating the inside and outside of the renal region and analyzing the entire area based on the 12 layers. The external (cortex) and internal (medullary) regions of the kidney have different structures and functions, and the TLCO method considers the peculiar structure of the kidney. In addition, it has been reported that the TLCO method is more stable than the ROI method because the entire renal region is stratified and analyzed simply by specifying the external and internal regions. However, images that are obtained clinically may be atrophied or deformed when the kidney is damaged and may show individual differences or contain cysts. As such, it is not always possible to divide kidneys evenly into multiple layers, resulting in unstable findings. Therefore, there is a desire for a fully automatic and stable method of kidney analysis. As a comprehensive quantification method for renal images, Kuo et al. proposed a method for estimating the eGFR at the time of examination using deep learning and ultrasonic images. In contrast, computer-aided diagnosis (CAD) studies on renal MRI and studies on transplanted kidneys have been reported. Khalifa et al. proposed a framework that estimates the rejection of transplanted kidneys using the time-series dynamic contrast-enhanced-magnetic resonance imaging (DCE-MRI) method. In this method, images obtained by time-series DCE-MRI are aligned, and the renal region is extracted using the level-set method. The alignment is then cor- rected, and the cortex is calculated and analyzed based on the brightness from the extracted renal region. Shehata et al. proposed a method for estimating the rejection of transplanted kidneys using deep learning. The DCE-MRI method requires a contrast medium and cannot be applied in cases of CKD. In addition, it is difficult to extract the cortex based on the brightness in impaired kidneys. In this study, we attempted to solve the above problems and develop a comprehensive evalu- ation method for renal MRI that can be applied as a clinical test. By taking advantage of the close relationship between renal morphology and renal function, the water images used in the Dixon method for evaluating the internal structure of the kidney were used to evaluate the target, and the correct label was eGFR, which is an index of renal function. Dixon techniques rely on the difference in resonance frequency between fat and water, and, therefore, fat-only, water-only, in-phase, and out-of-phase images are acquired. In the kidney, where the water-rich organ parenchyma is surrounded by fat, Dixon or similar fat-suppressed images clearly distinguish the renal parenchyma from the surrounding structure. We propose a method of extracting the renal region from magnetic resonance images with a U-net, converting the extracted renal region into a nonrigid body in the target mask, and then analyzing this region based on the TLCO of the target mask. Journal of Medical Imaging 024501-2 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Extraction of kidney area by U-net Nonrigid registration Estimation of eGFR by Regression Histogram U-net SVM Left-kidney Left-kidney Result of U-net Noise removal feature Original by 3D labeling Extraction of target kidney area and TLCO by manual (doctor) Original Kidney area TLCO Train mask Fig. 1 Proposed CAD system for estimation of the glomerular filtration rate slope from MRI using the Dixon method. 2 Material and Methods The proposed automated framework is shown in Fig. 1. The proposed framework uses the following four steps to process the MRI of the Dixon method: 1. Kidney area segmentation from the surrounding abdominal structures by U-net. 2. Noise removal by three-dimensional (3D) labeling. 3. Nonrigid registration of kidney area and target mask. 4. Calculation of the TLCO. 5. Estimation of the renal function (eGFR) by a regression support vector machine (SVM). In this paper, the proposed method is the automatic TLCO (A-TLCO) method that compen- sates for the weaknesses of the previously reported TLCO method and automates the measure- ment process. The details of the method are described below in order to distinguish the proposed method from the conventional manual TLCO method. 2.1 Kidney Area Is Segmented from the Surrounding Abdominal Structures by U-net Water images of the Dixon method used in this study have a clear renal region. In addition, this section describes a rough extraction of the renal region. Therefore, we have decided to use U-net, which is known to provide good results for the area extraction of medical images. Coronal sections of Dixon water images were used. There are three to six sliced images per subject, and the image size is 320 × 320. We used 1201 images of 174 cases. All images were created by a specialist as a renal segmentation image. Figure 2 shows the U-net used for renal region extrac- tion. The network discriminates between three classes: kidney, kidney boundaries, and other tissue. However, there is an imbalanced distribution of samples in the kidney boundaries class as compared with the other tissue class. We used a weighted cross-entropy loss in order to com- pensate for this imbalance and achieve more accurate learning when training the network. We used softmax with weighted cross-entropy loss for network output and true label comparison. 1 64 64 64 64 3 128 128 128 128 128 256 256 256 256 256 512 512 512 Fig. 2 The U-net architecture used for renal region extraction in this study. Journal of Medical Imaging 024501-3 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Fig. 3 Results of renal region extraction by U-net: (a) original, (b) mask, and (c) result. Cost minimization on 50 epochs was performed using the adaptive moment estimation optimizer with a learning rate of 0.0001. The training time for this network was ∼1h on a workstation with an NVIDIA TITAN RTX GPU × 2. A total of 1201 images were divided into 600 and 601 images, and classification was performed to extract the renal region. The renal function estima- tion experiment was then performed for all subjects. Figure 3 shows the extraction results obtained by U-net. Figure 3(a) is an input image, and Fig. 3(b) is a teacher image. Figure 3(c) shows the extraction result. It can be confirmed that there are few false positives and that the region close to the renal region can be extracted. 2.2 False-Detection Removal by 3D Labeling Although the renal region detected by U-net was highly accurate, false detection occurred. False positives are extracted for organs other than kidneys, as shown in Fig. 4(b). This may be because when U-net is trained with 64 × 64 patches, spatial information larger than the patch size is lost. To improve this problem, methods including detecting the location of the kidney by object detec- 14 15 tion, such as Faster R-CNN or YOLOv3, and applying semantic segmentation in a narrow region are considered. However, in this study, the boundary of the kidney by U-net is well dis- criminated, and there are few false positives for other organs, so there is little need to complicate the process. Therefore, we decided to perform 3D labeling for each subject and exclude areas Fig. 4 Kidney area obtained by U-net: (a) slice 1, (b) slice 2, and (c) slice 3. Result of the elimination of noise area by 3D labeling: (d) slice 1, (e) slice 2, and (f) slice 3. Journal of Medical Imaging 024501-4 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging other than those with a large area. The processed results are shown in Fig. 4. Figures 4(a)–4(c) show the extraction results by U-net. It can be confirmed that false positives occur in areas other than the renal region. Figures 4(d)–4(f) show the results of 3D labeling and exclusion of areas with a 3D area of 2500 pixels or less. This procedure was applied to all 1201 images, and it was confirmed that there were no cases of the renal region being accidentally excluded. 2.3 Nonrigid Registration of Kidney Area and Target Mask An effective method for estimating renal function is to analyze the structure of the kidney while 5,16 considering the kidney anatomically. However, time and effort are required for a specialist to manually extract the renal region. In addition, manual measurements will result in different results depending on the specialist. Therefore, we propose a method by which to automatically identify the positions of the cortex and medulla of all images by registration of all of the renal regions extracted by U-net to one target mask. The registration method is performed in two steps using a MATLAB function. The first step is a brightness-based affine transformation. The bright- ness-based registration method calculates the similarity between two images, repeats the affine transformation so that the similarity is high, and estimates the geometric transformation (trans- lation/rotation/scaling/shear) with the highest similarity. The processing results are shown in Fig. 5. Figure 5(a) is the original image. Figure 5(b) shows the results of extracting the renal region by U-net. Figure 5(c) is the target mask. Figure 5(c) is the renal region of a normal renal image extracted by a specialist. A typical patient has two kidneys. In this study, we analyze the kidney with a larger area. In Fig. 5(b), the left kidney is larger. As such, the left kidney is extracted and registered with the target mask. Figure 5(d) shows a pseudocolor image of the target mask and the initial position of the extracted kidney. The pseudocolor image shows the target mask in green, the extracted kidney in magenta, and the overlapping pixels of both in white. The result of the affine transformation is shown in Fig. 5(e). Next, the shape is finely 9,10 modified by registration using a brightness-based displacement field. In this paper, the regis- 1,2 tration is performed using the displacement field based on Thirion’s demons algorithm. The result of modification of the image shown in Fig. 5(e) using the displacement field is shown in Fig. 5(f). The fine shape has been modified to approach the target mask. Figure 5(g) shows the final registration of the image. Fig. 5 Result of rigid transform: (a) original image, (b) result by U-net, (c) target mask, (d) initial position, (e) affine transform, (f) displacement field, and (g) result of registration. Journal of Medical Imaging 024501-5 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Fig. 6 Calculation of the TLCO: (a) mask image, (b) seed image, (c) geodesic distance conver- sion, and (d) TLCO. 2.4 Calculation of TLCO In this study, the TLCO method is used for the analysis of renal function. The TLCO method specifies the inner (medullary side) and outer (cortical side) boundaries of the renal parenchyma and divides the renal region into 12 layers from the outside to the inside. The method for layering is not explained in the TLCO paper. Therefore, in this study, we calculated the TLCO using geodesic distance conversion. When geodesic distance conversion inputs a binary image of the renal region and a seed image, a distance conversion image is generated based on the seed image. Figure 6(a) shows a binary image of the renal region extracted by a specialist. Figure 6(b) shows a seed image made by a specialist. Figure 6(c) shows a pseudocolor image of the geodesic dis- tance conversion. However, the image shown in Fig. 6(c) was not obtained by dividing the renal region into 12 layers. Therefore, the result of geodesic distance conversion was normalized into 12 layers. Specifically, the number of layers was divided by the maximum value and then multi- plied by 12. The normalized geodesic distance conversion pseudocolor image is shown in Fig. 6(d). The target mask is very important because it directly affects the calculation of the TLCO method. In this study, we decided to use two types of images, a normal kidney image selected by a specialist and the kidney image with the largest region. Since the TLCO method divides the kidney into 12 layers, calculation with a small atrophied kidney is difficult. Therefore, a kidney image with a small area cannot be used as a target mask. Figure 7 shows the results of applying the TLCO method to MRI images of patients. Figure 7(a) is a normal image selected by a specialist. Figure 7(b) shows the results of the specialist extraction of the renal region from Fig. 7(a). Figure 7(c) is a pseudocolor image divided into 12 layers by the TLCO method. Figure 7(d) is an image of the kidney with the largest area. Figure 7(e) shows the result of the specialist extracting the renal region from Fig. 7(d). Figure 7(d) is a pseudocolor image divided into 12 layers by the TLCO method. 2.5 eGFR Used to Estimate Renal Function by RSVM In the TLCO method, we used 12 measurements, one for each layer, to estimate the cortico- medullary gradient of MR signal intensity based on the following definitions: cortical Journal of Medical Imaging 024501-6 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Fig. 7 TLCO of the target mask: (a) normal kidney, (b) mask, (c) TLCO, (d) large kidney, (e) mask, and (f) TLCO. (mean of the three superficial layers), medullary [mean of deep layers (layers 8 to 10)], and the gradient [linear fit to the data points (layers four to seven) when plotting MR signal intensity versus % of depth]. It has been reported that the gradient of the MR signal intensity correlates with the eGFR. As chronic damage progresses in the kidney, the difference between the outer and inner layers in the internal structure disappears. The nephrologist evaluates these changes from the images. Therefore, computing the gradient of MR signal intensity from the outer to inner layers, the fourth to seventh layers provides an index that replaces the human evaluation index. The process has been explained in previous reports, and this index is known to be useful in 5,16,18 the evaluation of kidney images. In this study, we decided to use machine learning to esti- mate renal function. At this time, eGFR is a continuous value. Therefore, eGFR is estimated using a regression support vector machine (RSVM). In addition, the TLCO method uses the average value of each layer. However, in this study, the histogram of each layer is used as a feature. The number of bins in the histogram is 50. The finally calculated features are the mean, kurtosis, skewness, median, a median of each TLCO (12 feats) and histogram (50 feats) of all renal regions, and the histogram of each TLCO (12 × 50 ¼ 600 feats) was calculated. In addi- tion, a noise-removed image was created using a median filter. The noise-removed image was created by applying the median filter once and twice. The feature was calculated from the noise- removed image by the same method. Therefore, the number of features calculated as a feature candidate is 1998. 3 Results 3.1 Patient Data and MRI Acquisition Protocol CKD patients with a sufficient number of visits for the treatment and evaluation of CKD (more than three times during a period of at least 1 year) were enrolled (n ¼ 165, 105 men and 55 women, aged from 19 to 87 years, and with a mean age of 60 15 years). The proposed frame- work has been tested on two-dimensional water images of the Dixon method data sets collected from 165 subjects. Magnetic resonance examinations were performed on a 3.0 T (Skyra, Siemens Healthcare, Erlangen, Germany) scanner using a spine coil and an 18-channel phased- array body coil. Coronal 3D T1-weighted volume-interpolated breath-hold examination Dixon images were generated from in-phase, out-of-phase, fat-only, and water-only images. The MRI Journal of Medical Imaging 024501-7 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging parameters were as follows: echo time (TE); delta TE (= 2.46 ms); delta (= 1.23); repetition time (= 5.35 ms); flip angle (= 10 deg); field of view (¼360 × 360 × 144 mm); recon matrix (= 320); and respiratory compensation (= breath-hold). During the observation period, blood and urine data were sampled every 3 to 6 months. The eGFR was calculated using the formula for Japanese 19 2 patients, i.e., eGFR (mL∕ min ∕1.73 m ) = 194 × Cr − 1.094 × age − 0.287 (× 0.739 for female cases). Since eGFR always decreases, we performed a linear approximation using blood test results from multiple time points and calculated the eGFR at the date of imaging. We also estimated the rate of decline in eGFR using the same linear approximation method and sub- sequently defined the “annual rate of decline in eGFR = eGFR slope” as an index of CKD prog- nosis. Urinary protein (UPro) levels were noted and corrected for urinary creatinine (urine protein: creatinine ratio, g:g). The Institutional Review Board of Saitama Medical University approved this research and its publication in accordance with the Declaration of Helsinki and the Japanese guidelines for clinical research (Approval Nos. 12065, 18107, 19048.01, 20107.01). 3.1.1 Kidney segmentation In this study, U-net is used to extract the renal region. To determine the hyperparameters of U-net, experiments were conducted with patch sizes of 32 × 32, 64 × 64,and 128 × 128 and two to four layers. In the experiment, 600 of the 1201 water images of the Dixon method were used for learning, and the extraction accuracy of 601 images was evaluated. The segmentation accuracy is evaluated using the Dice similarity coefficient, characterizing the agreement between the segmented and ground truth regions. The dynamic susceptibility contrast (DSC) is defined by the following expression: 2TP EQ-TARGET;temp:intralink-;sec3.1.1;116;446Dice coefficient ¼ ; 2TP þ FP þ FN where TP, FP, and FN denote the true positive, false positive, and false negative segmentation results, respectively. As illustrated in Table 1, the closer the DSC is to 1, the better the segmen- tation. To obtain the ground truth in our experiments, a medical specialist delineated the kidney borders. In addition, Table 2 shows the accuracy of the two-class (renal region and other tissues) segmentation, and Table 3 shows the accuracy of the three-class (renal region, renal region boun- dary, and other tissues) segmentation. In addition, erosion and dilation treatment was added as Table 1 Confusion matrix. Ground truth Kidney area Other area U-net result Kidney area TP FP Other areas FN TN Table 2 Accuracy of two-class segmentation result by U-net. The number of layers of U-net Two-class 2 (%) 3 (%) 4 (%) Patch size 32 × 32 77 79 80 64 × 64 79 82 82 128 × 128 77 82 83 Journal of Medical Imaging 024501-8 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Table 3 Accuracy of three-class segmentation result by U-net. The number of layers of U-net Three-class 2 (%) 3 (%) 4 (%) Patch size 32 × 32 82 84 85 64 × 64 84 87 86 128 × 128 82 86 87 needed. The best accuracy was 87%. The most accurate results were for the three-class mask (renal region, renal region boundary, and other tissues), three layers of U-net, and a patch size of 64 × 64. Similar results were obtained for four layers of U-net and a patch size of 128 × 128. In this study, we decided to use three layers of 64 × 64 in consideration of learning time. 3.1.2 Comparison of the TLCO and A-TLCO methods In this study, we estimate the eGFR by multiple regression analysis using the average value of each layer of the TLCO calculated from water images of the Dixon method for comparison with the A-TLCO method. The estimated results by the TLCO method are shown in Table 4. Multiple water images of the Dixon method were obtained for each subject. In this study, we analyze the kidney with the largest area for each subject. UPro is known to correlate with renal function in urine and blood tests. The correlation coefficient is shown in Table 4. Upro had a negative cor- relation of −0.32. On the other hand, the result of manually creating TLCO and estimation by multiple regression analysis was 0.49. The TLCO method results correlate better with the eGFR as compared with the UPro. On the other hand, the correlation coefficient estimated using multi- ple regression analysis calculated by the A-TLCO was 0.46. Therefore, it was clarified that the proposed method enables TLCO analysis without the need for a specialist to select the renal region. 3.1.3 Estimation of renal function by RSVM The accuracy of the A-TLCO method was evaluated using an RSVM and a regression random forest (R-RF). The root mean square error (RMSE) and correlation coefficient were used for evalu- ation. The RMSE was calculated as the accuracy evaluation of continuous values, as follows: vﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ u X t 2 EQ-TARGET;temp:intralink-;sec3.1.3;116;297RMSE ¼ ðF − T Þ ; i i i¼1 where N is the number of estimated values, F is the estimated value, and T is the true value. In addition, we decided to calculate the correlation coefficient to verify the clinical significance of the estimation 1 F − μ T − μ i F i T EQ-TARGET;temp:intralink-;sec3.1.3;116;204CorrelationðF; TÞ¼ ; N − 1 σ σ F T i¼1 Table 4 Correlation of eGFR by multiple regression analysis. UPro Manual Affine Affine&Int Accuracy of registration —— 0.9 0.88 Correlation −0.32 0.49 0.46 0.42 Journal of Medical Imaging 024501-9 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging where μ is the average and σ is the standard deviation. A-TLCO registration was performed with a target mask. In this case, the accuracy may differ depending on the target mask. The A-TLCO divides the renal region into 12 layers. Since dividing the kidney into 12 layers is difficult if the kidney is atrophied, we decided to use two types of kidneys: kidneys in good condition that are closest to normal and kidneys that have the largest area among the normal cases as selected by a specialist. In addition, for the registration method, we compared the affine transformation and the result of applying the displacement field-based registration method in addition to the affine trans- formation. Feature selection was performed using the forward step-wise method. Subject-based cross-validation was used for accuracy evaluation. Table 5 shows the results of eGFR at the time of MRI imaging. The maximum value of the eGFR used in this study is 113, and the minimum value of the eGFR is 9.4. The best estimation correlation coefficient of eGFR was obtained by using a kidney with a large area as the target mask, applying affine and displacement field-based transformations for registration, and using RSVM, with an RMSE of 11.99 and a correlation coefficient of 0.83. An expert manually calculated 165 TLCO and estimated these values using RSVM results with an RMSE of 15.07 and a correlation coefficient of 0.69. The RSVM models were created using MATLAB 2019a in the experiment. The training option and parameters used for training were as follows: kernel = rbf, box constraint = iqrðYÞ∕1.349, and ε = iqrðYÞ∕1.349. Here, iqrðYÞ is the interquartile range of response variable Y. The kernel width parameter σ was determined by a heuristic subsampling procedure. We compared the SVM with the R-RF. The number of trees was set to 300, and the parameters were used by default. The results obtained using the R-RF indicated that the RMSE of the eGFR was 14.48 and that the correlation coef- ficient was 0.76. In addition, we compared the results using the deep learning of GoogLeNet. To use GoogLeNet in this study, the last fully connected layer, softmax, and the classification output of the network were deleted, and the fully connected layer and regression layer of output 1 were added. GoogLeNet uses the result of learning with ImageNet as transfer learning. The training option and parameters used for training were as follows: initial learning rate = 0.005, Max Epochs = 100, l2reg = 0.0001, and Mini Batch Size = 64. Subject-based cross-validation was used for accuracy evaluation. The results obtained using GoogLeNet indicated that the RMSE of the eGFR was 21 and the correlation coefficient was 0.23. We also compared the method with deep learning and SVM (GoogLeNet + SVM), which has recently been the focus of attention. Table 5 Accuracy of eGFR by machine learning. eGFR (present) Registration The number Discriminator Mask type method of features RMSE Corr Regression-SVM Normal kidney Affine 12.17 0.83 21 Affine & Disp 14.06 0.74 13 Large kidney Affine 12.89 0.8 19 Affine & Disp 11.99 0.83 31 R-RF Normal kidney Affine 14.96 0.71 7 Affine & Disp 15.47 0.67 8 Large kidney Affine 14.45 0.76 8 Affine & Disp 14.48 0.76 8 Regression-SVM Manual extraction 15.07 0.69 11 GoogLeNet 21 0.23 — GoogLeNet + SVM 21 −0.01 — Note: Bold value is the discriminator with the highest correlation coefficient. Journal of Medical Imaging 024501-10 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Table 6 Estimation accuracy of eGFR slope by machine learning. eGFR slope Registration The number Discriminator Mask type method of features RMSE Corr Regression- Normal kidney Affine 4.87 0.4 8 SVM Affine & Int 4.8 0.48 9 Large kidney Affine 4.86 0.42 5 Affine & Int 4.8 0.5 14 R-RF Normal kidney Affine 4.68 0.44 5 Affine & Int 4.58 0.5 8 Large kidney Affine 5.00 0.28 4 Affine & Int 4.73 0.42 5 Regression- Manual extraction 4.82 0.45 7 SVM GoogLeNet 5.35 −0.09 — GoogLeNet + SVM 5.25 0.11 — Note: Bold value is the discriminator with the highest correlation coefficient. With the pretrained GoogLeNet, the high-dimensional features that occur after obtaining the deep features of the MRI images are reduced with the neighbor component analysis (NCA) 21,22 method. The accuracy was evaluated by SVM using the features obtained by NCA. The results obtained using GoogLeNet + SVM indicated that the RMSE of the eGFR was 21 and that the correlation coefficient was −0.01. Table 6 shows the estimation results for eGFR slope, which indicates the prognosis of renal function. The maximum value of eGFR slope used in this study is 50.5, and the minimum value is −27.6. The best estimation correlation coefficient of the eGFR slope was obtained by using a kidney with a large area as the target mask, applying affine and displacement field-based transformations for registration, and using an RSVM with an RMSE of 4.8 and a correlation coefficient of 0.5. A similar correlation coefficient was obtained when using an RSVM with an RMSE of 4.58, and a correlation coefficient of 0.5. An expert manually calculated 165 TLCO and estimated the TLCO using RSVM results with an RMSE of 4.82 and a correlation coefficient of 0.45. GoogLeNet was calculated to obtain an RMSE of 5.35 and a correlation coefficient of −0.09. GoogLeNet + SVM was calculated to obtain an RMSE of 5.25 and a correlation coefficient of −0.11. 3.2 Selected Features of the Proposed Method In this study, 1998 features were calculated, and feature selection was performed using the for- ward step-wise method. The eGFR estimation required an average of 14 features. The least-cases scenario involved seven features, and the most-cases scenario involved 31 features. Table 5 indicates the number of features required for convergence. Table 7 shows the features selected for estimating eGFR (RSVM + Large kidney + Affine & Int). The histograms calculated from the entire mask were selected from 7 out of 31 features. However, most of the histograms were created for each A-TLCO layer and 24 of the 31 features. Figure 8 shows which layer was selected in the TLCO layer-by-layer histogram. At this time, layer 1 is the outermost layer, and layer 12 is the innermost layer. Layers 4 to 6 are the boundaries between the cortex and the medulla. Although the features are selected from the histogram of each layer, relatively many bins of the layers corresponding to the inside and outside are selected. This may indicate that changes in the cortex and medulla are effective. In addition, many features of the image from which noise was removed by applying a median filter were selected. Journal of Medical Imaging 024501-11 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Table 7 Features selected for eGFR estimation. Feature name Number of features Mean, kurtosis, skewness, median (4 × 3 = 12 feats) 0 Histogram of kidney (50 feats) 1 Histogram of kidney of median filter one time (50 feats) 1 Histogram of kidney of median filter two times (50 feats) 5 Median of each layer of A-TLCO (12 × 3 = 36 feats) 0 Histogram of each layer of A-TLCO (600 feats) 7 Histogram of each layer of A-TLCO of median filter one time (600 feats) 10 Histogram of each layer of A-TLCO of median filter two times (600 feats) 7 Fig. 8 A-TLCO histogram features selected by eGFR estimation. Table 8 Features selected for eGFR slope estimation. Feature name Number of features Mean, kurtosis, skewness, median (4 × 3 = 12 feats) 1 Histogram of kidney (50 feats) 0 Histogram of kidney of median filter one time (50 feats) 1 Histogram of kidney of median filter two times (50 feats) 0 Median of each layer of A-TLCO (12 × 3 = 36 feats) 0 Histogram of each layer of A-TLCO (600 feats) 4 Histogram of each layer of A-TLCO of median filter one time (600 feats) 0 Histogram of each layer of A-TLCO of median filter two times (600 feats) 8 Table 7 shows the number of features required for convergence of the eGFR. Table 8 shows the features selected for estimating the eGFR slope (RSVM + Large kidney + Affine & Int). The mean, kurtosis, skewness, median, and histogram calculated from the entire mask were selected from 2 of 14 features. However, most of the histograms were created for each A-TLCO layer and 12 of the 14 features. Figure 9 shows which layer was selected in the TLCO layer-by-layer histogram. As few as 12 features are selected, and relatively many bins of the layers corresponding Journal of Medical Imaging 024501-12 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging Fig. 9 A-TLCO histogram features selected by slope estimation. to the inside and outside are selected. In addition, many features of the image from which noise was removed by applying a median filter were selected. 4 Discussions 4.1 Performance of the Proposed Method In this study, renal function was estimated using water images of the Dixon method. The eGFR accuracy was 0.38 when estimated by multiple regression analysis using the average value for each layer. On the other hand, when a normal kidney with a large area was used as the renal region and affine transformation and displacement field registration were used for registration, the RMSE was 11.99 and the correlation coefficient was 0.83, which is a large improvement. Large kidneys tended to have better RMSE and correlation coefficients than normal kidneys. For comparison, features were calculated and evaluated using the TLCO manually extracted by a specialist. The TLCO method has an RMSE of 15.07 and a correlation coefficient of 0.69. Better results were obtained with the A-TLCO method than with the TLCO method. It is thought that the reason for this is that the value provided by the TLCO method varies because the renal region is manually divided into layers and dividing small kidneys, such as atrophic kidneys, into 12 layers is difficult. The results obtained by GoogLeNet had an RMSE of 21 and a correlation coefficient of 0.23. Using GoogLeNet + SVM did not provide good results, giving an RMSE of 21 and a correlation coefficient of −0.01. It became clear that estimation by simple deep learning was difficult. This probably indicates the small number of data used for learning as well as the difficult estimation of information that specialists cannot recognize from images, using deep learning. However, using the A-TLCO method, we found that eGFR could be esti- mated with high accuracy from water images of the Dixon method. A similar experiment was performed on the eGFR slope. The eGFR slope was calculated using a normal kidney with a large area as a renal region, and affine transformation and dis- placement field registration were used for registration. The RMSE was 4.8, and the correlation coefficient was 0.5. In the estimation of the eGFR slope, a correlation coefficient of 0.5 was also obtained using R-RF. However, the estimated eGFR slope tended to have a significantly lower correlation coefficient than the eGFR estimation. The TLCO manually extracted by a specialist was also evaluated using the same method. The RMSE was 4.82, and the correlation coefficient was 0.45. Better results were obtained with the A-TLCO method. The results obtained by GoogLeNet had an RMSE of 5.35 and a correlation coefficient of −0.09. Using GoogLeNet +SVM did not provide good results, giving an RMSE of 5.25 and a correlation coefficient of 0.11. It was again clear that estimation by simple deep learning was difficult. However, it was found that the prognosis of renal function can be estimated to some extent using the A-TLCO method, even using water images of the Dixon method. In order to improve the accu- racy of eGFR slope estimation, it is necessary to consider different MRI images, such as the T2* map generated by BOLD MRI. Journal of Medical Imaging 024501-13 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging 4.2 Adequacy for Clinical Application of Selected Features of the Proposed Method Nephrologists have empirically predicted renal function from renal morphology. In diagnosing a case of renal dysfunction of unknown course, acute kidney injury and chronic kidney disease are diagnosed from blood tests and an imaging study including MRI, and an examination/treatment policy is determined. For example, the discrepancy between the predicted eGFR from the images and the actual eGFR is known to be large in the case of acute kidney injury, while in the case of chronic kidney injury, the discrepancy is small. In addition, MRI is used as a reference for predicting renal function recovery by treating patients with acute kidney injury. However, neph- rologists empirically predict renal function based on the size of the kidney, surface irregularities, and the clarity of the boundary between the cortex and medulla, and quantitative evaluation is difficult. On the other hand, this study is expected to allow quantitative assessment and repro- ducibility to diagnostic imaging by MRI. The proposed method is a framework for quantifying the state of the kidney and can be applied to various types of MRI, such as T2* mapping. Therefore, the proposed method is expected to become a new imaging method for evaluating various physiological and pathological aspects of the kidney. 5 Conclusion We proposed the A-TLCO method to estimate renal function from water images obtained using the Dixon method. The proposed method afforded an RMSE of 11.99 and a correlation coefficient of 0.83 for eGFR estimation, an RMSE of 4.8 and a correlation coefficient of 0.5 for eGFR slope, which is an indicator of CKD prognosis. These results confirmed that the reliability was high and the accuracy was dramatically improved compared with the conventional TLCO method. The de facto standard for MRI measurement is the ROI method. Since the kidneys are con- centrically arranged with functional and anatomical distinctions, it has been reported that the TLCO method provides more information than does the ROI method. In this paper, by adding auto-annotation and registration processing using the mask, as well as devising the features used in the estimation, we have been able to greatly improve the accuracy of the estimation using the features obtained by the conventional TLCO method. MRI methods that can acquire the status of hypoxia and fibrosis have been reported, just as water imaging can image the internal structure of the kidney. Unlike x-ray computerized tomog- raphy and ultrasonography, MRI can also provide the physiological and functional aspects of the kidney along with positional information. As such, the proposed method of comprehensive MRI quantification is expected to afford an important imaging analysis method that can provide useful information to clinicians. Disclosures The authors declare no conflicts of interest associated with this manuscript. Acknowledgments The present research was supported by AMED through Grant No. JP19ek0210122h0001 and by JSPS KAKENHI Grant No. 21K19926. References 1. Japanese Society of Nephrology, “Essential points from evidence-based clinical practice guidelines for chronic kidney disease 2018,” Clin. Exp. Nephrol. 23(1), 1–15 (2019). 2. A. C. Webster et al., “Chronic kidney disease,” Lancet 389(10075), 1238–1252 (2017). 3. T. K. Chen, D. H. Knicely, and M. E. Grams, “Chronic kidney disease diagnosis and management: a review,” J. Am. Med. Assoc. 322(13), 1294–1304 (2019). Journal of Medical Imaging 024501-14 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging 4. R. Simms and S. Sourbron, “Recent findings on the clinical utility of renal magnetic resonance imaging biomarkers,” Nephrol. Dial. Transplant. 35(6), 915–919 (2020). 5. M. Pruijm et al., “Reduced cortical oxygenation predicts a progressive decline of renal function in patients with chronic kidney disease,” Kidney Int. 93(4), 932–940 (2018). 6. K. Sugiyama et al., “Reduced oxygenation but not fibrosis defined by functional magnetic resonance imaging predicts the long-term progression of chronic kidney disease,” Nephrol. Dial. Transplant. 35, 964–970 (2018). 7. T. Inoue et al., “Noninvasive evaluation of kidney hypoxia and fibrosis using magnetic resonance imaging,” J. Am. Soc. Nephrol. 22(8), 1429–1434 (2011). 8. L. P. Li et al., “Renal BOLD MRI in patients with chronic kidney disease: comparison of the semi-automated twelve layer concentric objects (TLCO) and manual ROI methods,” MAGMA 33(1), 113–120 (2020). 9. C. C. Kuo et al., “Automation of the kidney function prediction and classification through ultrasound-based kidney imaging using deep learning,” NPJ Digital Med. 2, 29 (2019). 10. F. Khalifa et al., “Dynamic contrast-enhanced MRI-based early detection of acute renal transplant rejection,” IEEE Trans. Med. Imaging 32(10), 1910–1927 (2013). 11. M. Shehata et al., “Computer-aided diagnostic system for early detection of acute renal transplant rejection using diffusion-weighted MRI,” IEEE Trans. Biomed. Eng. 66(2), 539–552 (2019). 12. J. Ma, “Breath-hold water and fat imaging using a dual-echo two-point Dixon technique with an efficient and robust phase-correction algorithm,” Magn. Reson. Med. 52(2), 415–419 (2004). 13. O. Ronneberger, P. Fischer, and T. Brox. “U-Net: convolutional networks for biomedical image segmentation,” Lect. Notes Comput. Sci. 9351, 234–241 (2015). 14. K. He et al., “Mask R-CNN,” in IEEE Int. Conf. Comput. Vision, pp. 2980–2988 (2017). 15. J. Redmon and A. Farhadi. “YOLOv3: an incremental improvement,” arXiv:1804.02767 (2018). 16. B. Milani et al., “Reduction of cortical oxygenation in chronic kidney disease: evidence obtained with a new analysis method of blood oxygenation level-dependent magnetic resonance imaging,” Nephrol. Dial. Transplant. 32(12), 2097–2105 (2017). 17. P. Soille, Morphological Image Analysis: Principles and Applications, 2nd ed., pp. 219– 221, Springer-Verlag, Secaucus, New Jersey (2003). 18. T. Inoue et al., “Comparison of multiparametric magnetic resonance imaging sequences with laboratory parameters for prognosticating renal function in chronic kidney disease,” Sci. Rep. 11(1), 22129 (2021). 19. S. Matsuo et al., “Collaborators developing the Japanese equation for estimated GFR. Revised equations for estimated GFR from serum creatinine in Japan,” Am. J. Kidney Dis. 53(6), 982–992 (2009). 20. J. Deng et al., “ImageNet: a large-scale hierarchical image database,” in IEEE Conf. Comput. Vision and Pattern Recognit., Miami, Florida, pp. 248–255 (2009). 21. A. Çınar and S. A. Tuncer, “Classification of lymphocytes, monocytes, eosinophils, and neutrophils on white blood cells using hybrid Alexnet-GoogleNet-SVM,” SN Appl. Sci. 3(4), 503 (2021). 22. E. Basar ¸ an, Z. Cömert, and Y. Çelik, “Neighbourhood component analysis and deep feature-based diagnosis model for middle ear otoscope images,” Neural Comput. Appl. 1–12 (2022). Masahiro Ishikawa is an assistant professor at Saitama Medical University. He received his MS degree in engineering from Niigata University in 2004 and received his PhD in engineering from Niigata University in 2007. His current research interests include medical image processing, pathology image recognition, and computer-aided diagnosis. Tsutomu Inoue MD, PhD is a nephrologist. He is currently an associate professor and chief of research in the Department of Nephrology, Faculty of Medicine, Saitama Medical University. His research focuses on analytical methods and clinical applications of renal MRI as well as the molecular mechanisms underlying the progression of chronic kidney disease. He provided the Journal of Medical Imaging 024501-15 Mar∕Apr 2022 Vol. 9(2) Ishikawa et al.: Framework for estimating renal function using magnetic resonance imaging discussion of the validity and clinical significance of the results obtained by the method proposed in this paper. Eito Kozawa MD, PhD is a clinical and diagnostic radiologist. He is currently a professor and chief administrator of the Department of Radiology, Faculty of Medicine, Saitama Medical University. He is involved in education in his area of expertise at the university and university hospital. His clinical and research interests include MRI imaging and analysis, and he played the role of expert in relation to the renal MRI imaging and analysis techniques in this study. Hirokazu Okada was graduated from Keio University in 1987. He received his MD/PhD degrees from Keio University, and undertook a research fellowship at the Renal Division, University of Pennsylvania. He joined the Department of Nephrology, Saitama Medical University in 1996. He has been managing the department as professor since 2013, serving as the deputy-dean since 2018 and as the vice-president since 2020. He provided the discussion of the results in this paper. Naoki Kobayashi received his PhD from Niigata University, Niigata, Japan, in 2000. He worked for Laboratories of Nippon Telephone and Telegram co. from 1981 to 2008. He is a professor in School of Biomedical Engineering, Faculty of Health and Medical Care of Saitama Medical University since 2008. His research interest is medical image processing and bio-signal processing. Journal of Medical Imaging 024501-16 Mar∕Apr 2022 Vol. 9(2)
Journal of Medical Imaging – SPIE
Published: Mar 1, 2022
Keywords: quantitative estimated glomerular filtration rate; magnetic resonance imaging; kidney
Access the full text.
Sign up today, get DeepDyve free for 14 days.