An automatic segmentation framework for computer-assisted renal scintigraphy procedure

Abstract

One of the techniques for achieving unique and reliable information in medicine is renal scintigraphy. A key step for quantitative renal scintigraphy is segmentation of the kidneys. Here, an automatic segmentation framework was proposed for computer-aided renal scintigraphy procedures. To extract kidney boundary in dynamic renal scintigraphic images, a multi-step approach was proposed. This technique is featured with key steps, namely, localization and segmentation. At first, the ROI of each kidney was estimated using Otsu’s thresholding, anatomical constraint, and integral projection, which is done in an automatic process. Afterwards, the ROI obtained for the kidneys was used as the initial contours to create the final counter of kidneys using geometric active contours. At this step and for the segmentation, an improved variational level set was utilized through Mumford-Shah formulation. Using e.cam gamma camera system (SIEMENS), 30 data sets were used to assess the proposed method. By comparing the manually outlined borders, the performance of the proposed method was shown. Different measures were used to examine the performance. It was found that the proposed segmentation method managed to extract the kidney boundary in renal scintigraphic images. The proposed technique achieved a sensitivity of 95.15% and a specificity of 95.33%. In addition, the section under the curve in the ROC analysis was equal to 0.974. The proposed technique successfully segmented the renal contour in dynamic renal scintigraphy. Using all the data sets, a correct segmentation of the kidney was performed. In addition, the technique was successful with noisy and low-resolution images and challenging cases with close interfering activities such as liver and spleen activities.

1. Introduction

Blood wastes such as urea and salt are removed and filtered by the kidneys and excreted from the body through urine. Functional renal scintigraphy, also known as “renal scanning,” is an effective and relatively inexpensive procedure which could provide reliable information about perfusion, parenchyma, and excretory function of kidneys and follow the urinary draining route, including ureters and urinary bladder. This procedure is carried out by small amounts of compound containing radioisotope and chemical agent, which is externally traced with a gamma camera and saved in a computer [1]. As a result, renal function could be quantified with specific software to delineate renal outline and draw time-activity curve of kidneys, named renogram. In this regard, precise diagnosis and prognosis, renogram is highly desired. Therefore, kidney problems could be diagnosed in its earlier stages and followed up during time without invasive techniques or surgery [2].

Renal scintigraphy can provide unique information to the physician that is often unachievable using other imaging procedures [3]. One of the complicated procedures is renal segmentation, and the difficulty depends on image modality [4]. Several authors have developed different techniques for Kidney segmentation from different imaging modalities like as CT, MRI, US, etc. In [5], a 3D segmentation method based on a geometric deformable model directed by a special stochastic speed function was proposed for kidney borders segmentation in CT images. Authors in [4] introduced a comprehensive review of the latest techniques of renal segmentation applied to CT imaging modality. In [6], the level set algorithm using prior shape knowledge was developed for contouring kidney parenchyma in the MRI data set. The approach in [7] was based on applying an adjusted distance regularized level set that utilizes local features to segment kidney in ultrasound images.

As far as the authors know, segmentation of kidney, in dynamic renal scintigraphy dataset in particular is an area that has not been evaluated thoroughly. In [8], Sobel operator followed by Unsharp masking and image histograms was used to segment kidneys from the background. In [9], a hybrid technique based on anatomical knowledge and active shape models (ASMs) was presented for kidney segmentation. In [10] and [11], an automated system to define intended renal regions was introduced based on HOG3D spatiotemporal descriptor by taking into account the advantage of the fast marching method (FMM). The approach in [12] kidney was segmented the kidney in dynamic renal scintigraphic images by combination the advantages of FMM in conjunction with Harris corner descriptor. The level sets method uses an Eikonal equation was utilized in [13] to trace kidneys border in gamma camera images.

Segmentation of the kidneys from SPECT or SPECT/CT images is important in modern quantitative scintigraphy procedures, such as measurement of glomerular filtration rate (GFR). Also, accurate knowledge of the border of kidney is very important in order to developing and training a deeplearning-based kidney segmentation for renal scintigraphy analysis system [14] and [15]. Manual segmentation in the field of segmentation is a labor-intensive and tedious task. It is featured with impracticable and inexact calculation results, so that when it is done by an expert; the result varies up to 20% [16]. Moreover, conventional techniques like thresholding used in many commercial platforms, are not enough for precise extracting of kidneys in renal scintigraphic data sets. Given this, developing automatic segmentation techniques for quantitative renal scintigraphy are required. These techniques are featured with notable accuracy and a limited user interaction. They are suitable for complete utilization of medical data.

In [17], a hybrid method based on a variational level set method was introduced for segmentation of left ventricle (LV) in myocardial perfusion scintigraphy (MPS) volumetric. Here, given the anatomical constraint of kidneys and a new variational level set formulation, a multi-step approach was proposed for fully automatic extracting of kidneys border in dynamic scintigraphy images. Section 2 represents the data set and proposed techniques, followed by Section 3 on experimental results. Section 4, clarifies the practical features of the algorithms, including advantages and the limitations. Eventually, Section 5 gives the conclusion and recommendations for future words.

Nowadays, the segmentation is performed semi-automatically in clinical practice. Specialists select ROI by drawing contours around kidneys, it is a challenging issue due to human intervention and time-consuming. There are also some limitations in a clinical method for problematical cases which can make specialists to be confused.

This study is an attempt to propose an efficient method that has the potential to be used for segmentation of kidneys based on renal scintigraphy data. The method is designed to meet the needs in clinical setting and overcome some of the limitations of available renography software in kidney segmentation. In this regard, the proposed method can minimize human interference for drawing kidney contours and can reduce inter-observer and even intra-observer variability. The proposed techniques present acceptable results in problematic cases, but the main limitation of the current study is in few cases due to very high background noise because of the chronic renal failure. To evaluate validity of the proposed approach, the obtained automatically recognized boundaries were compared to the boundaries that have been manually identified by specialists.

2. Methodology

This section describes, in details, the data set and proposed technique. It contains two parts. Initially, data acquisition and processing protocol are introduced, followed by introducing the proposed method for extracting kidneys border in renal scintigraphic images.

2.1. Acquisition and processing protocol

Totally, 18 female and 12 male patients with explicit or suspected obstructive uropathy took part in this research, ranging in age from 45 to 78 years. (mean age: 58.30 ± 9.02). Images were acquired in posterior view while the patient lays on supine position. Imaging was performed immediately after intravenous injection of 10–15 mCi (0.370–0.555 GBq) of Tc-99 m DTPA. Data in renal scintigraphy has 40 frames for each patient. Dynamic images were obtained in one frame per minute, and the process takes 40 min for each patient (A total of 40 frames per patient). Hence, for thirty patients, all of 40 frames were used in this study to segment kidneys and then obtain renogram curves (30 patients * 40 frames = 1200 images in total). Frames by e-cam gamma camera (SIEMENS-Germany) featured with a low-energy high resolution parallel-hole collimator and images were formed with 64 × 64 matrices for semi-quantitative study, region of interest (ROI) was drawn out at peak of activity (about 2–3 min) around kidneys as well as the other semilunar perirenal ROI for background correction.

2.2. Proposed method of kidney segmentation

To make the calculation of renal functions’ parameters feasible, a multi-step approach was proposed for segmentation and track the kidneys in dynamic renal scintigraphic data set. The main steps of the techniques proposed here include (1) ROI localization and (2) segmentation. The time-activity curves registered at each kidney, following the segmentation of kidneys in all images, are plotted. Based on the renogram curves, we can determine the differential function. The block diagram of the techniques is illustrated in Fig. 1.

Fig. 1.

Proposed segmentation framework for quantitative renal scintigraphy. The proposed method has four stages. At first, ROI localization is determined for right and left kidneys in each frame. Then, in the segmentation part, kidney contours are determined. Afterward, the activity of each kidney will be obtained in each frame. Finally, the renogram curves are drawn

After drawing ROI around each kidney, time-activity curve is generated. It shows initial flow and cortical update phase (upslope of the curve) which contains extraction and wasteful of trace flow collecting system (downslope of the curve). In this patient, third phase of time-activity curve (TAC) is somewhat slower than usual.

2.2.1. ROI localization

In addition to the kidney, other organs like bladder, liver, and spleen can be also visible in renal scintigraphic images. This step is aimed at segmenting the renal scintigraphy images into distinct boxes, so that each box only has one kidney. In this way, ROI is obtained for each kidney. Therefore, the computational load is reduced and, as a result, the speed of the process is increased. We propose a hybrid algorithm based on anatomical constraint and integral projection [18] to automatically estimate the boundary of each kidney.

In this regard, an image that contains kidneys and bladder is binarized at first using an adaptive thresholding like Otsu’s thresholding [19]. The technique was successful in determining the required threshold values automatically (Fig. 2b). Then a process of horizontal and vertical integral projection that searches for the lines that are separated from the object border is provided.

Fig. 2.

An example for achieving ROI localization left and right kidneys. a An original renal scintigraphic image. b Binarized image is obtained by the Otsu’s thresholding. c An original image is overlaid with separation refined lines

Due to anatomical constraints and, given that we can separate the kidneys from the bladder using a straight line, the x-axis projection histogram along that line is minimized. Separation of each individual kidney can be done using the projection histogram in the Y-axis direction. Mathematical formulation of integral projection in horizontal and vertical direction is $\text{H}(x)=\sum _{y}I(x,y)\text{and}V(y)=\sum _{x}I(x,y)$, respectively. Where the I be the binarized image.

We estimated the distance of horizontal line, which is separating the kidneys and bladder, to the top of the original image is two third the length of image and the distance of vertical line, which is separating kidneys, to the other two sides of the original images are equal. Thus, we have a closed box that perfectly fixes the 48 × 32 areas and, it constructs ROI for each kidney in an automatic manner. The position of center of these rectangular is (24, 16) and (24, 48) for left and right kidneys, respectively. The result is given in Fig. 2c.

2.2.2. Segmentation

The initial closed curves were used in “ROI localization” to extract the final boundary using the variational level set method. The results of the introduced technique were utilized through connected component analysis (CCA) with 8-connectivity [20] to obtain the refined contour of kidneys (Fig. 1).

The active contour model introduced by [21], is extensively studied and utilized for image analysis. Active contours refer to curves that are deformed in digital images and recover object shapes. Based on their representation and implementation, these contours are categorized as parametric or [22] geometric [23] models. Based on the curve evolution theory that was applied through the level set algorithms, geometric active contours are given implicitly [24]. In the field of segmentation, the level set method is another approach to solve energy minimization problem. Level set method using edges is obtained based on the derivatives of image intensities.

Though the low-quality scintigraphic images, the renal scintigraphy data have a low signal to noise and the kidney can be obtained only when an external model is available. Therefore, these methods are not capable of extracting the kidneys’ contour accurately. To deal with this limitation, a variational level set method was used [25]. This method employs an energy functional that consists of surface tension and bulk energies. It also mixes theoretical variational formulation and the level set method. This approach has a good robustness to noise, which is an important challenge for most of the traditional techniques.

A new variational level set method was employed using Mumford–Shah formulation [26] for segmentation. Given our earlier studies [27], contours with or without a gradient can be extracted by this model. Therefore, it is possible to segment kidneys with smooth borders or even borders with discontinuity problems. Furthermore, the model also has a good performance with the initial contours, while the conventional level set method can be affected by the position of the initial contours. Thus, the proposed method is capable of estimating the border of the kidney accurately, regardless of the starting position of the initial contour contained in the original scintigraphic image. The formulation of the model is as follows.

$${\inf }_{\left(c_1,c_2,C\right)}E=\mu \text{Length}(C)+\nu Area(Inside(C))+E_{MV},$$ (1)

with

$$E_{MV}={\lambda }_1\underset{\mathrm{inside}(C)}{\int }{\left(u_0(x,y)-c_1\right)}^{2}dxdy+{\lambda }_2\underset{\text{outside}(C)}{\int }{\left(u_0(x,y)-c_2\right)}^{2}dxdy,$$ (2)

where u₀ is the original image, and C is a parameterized curve, and C_i is the means of u₀ outside and inside C, and μ ≥ 0, ν ≥ 0, and λ_i, ≻ 0 are fixed parameters. The level set function obtained in [26] is:

$$\{\begin{array}{c}\frac{\partial \phi }{\partial t}={\delta }_{\epsilon }(\phi )\left[\mu \cdot div\left(\frac{\nabla \phi }{|\nabla \phi |}\right)-v-{\lambda }_1{\left(u_0-c_1\right)}^2+{\lambda }_2{\left(u_0-c_2\right)}^2\right]\\ \phi (x,y,0)={\phi }_0(x,y)in\text{Ω},\\ \frac{{\delta }_{\epsilon }(\phi )\partial \phi }{|\nabla \phi |\partial n}=0on\partial \text{Ω},\end{array}$$ (3)

where n represents the exterior to the boundary ∂Ω, and ∂φ/∂η refers to normal derivative of φ at the boundary and δ_ε represents the Dirac delta function. To handle this partial differential equation, H_ε(φ) and δ_ε(φ) need to be regularized at first.

$$\{\begin{array}{c}{H}_{\alpha }(\phi )=\frac{1}{2}+\frac{1}{\pi }\arctan \left(\frac{\phi }{s}\right),\\ {\delta }_s(\phi )=\frac{1}{\pi }\cdot \frac{\epsilon }{{\epsilon }^2+{\phi }^2}.\end{array}$$ (4)

Clearly, ε → 0, H_ε(φ) converges to H(φ) and δ_ε(φ) converges to δ(φ)

$$H(\phi )=\{\begin{array}{ll}1& if\phi \ge 0\\ 0& if\phi <0\end{array},\delta (\phi )=\frac{d}{d\phi }H(\phi ).$$ (5)

3. Experiments

The performance of the proposed technique is illustrated through comparing the manually outlined borders. To this end, 30 different data sets were used. Implementations were carried out using MATLAB 2019a (www.mathworks.com) (Intel Core i5-8400, with 16 GB-DDR4, Microsoft Windows 10-(64-bit)).

As mentioned, the ROI was performed at first. Then, it was used for segmentation of kidney using the variational level set method. The results of the proposed technique were utilized through CCA to obtain the contour of the kidneys in scintigraphy images. Figures 3 and 4 give the final boundaries of the kidney following using the variational level set.

Fig. 3.

Renal scintigraphy data from one adult case in some frames, from top to bottom, a original data, b extraction of the kidney contours, c kidney areas for counting activities, and d final boundaries of the kidney

Fig. 4.

Scintigraphic data from seven adult cases, from top to bottom, a original data, b extraction of kidney contours, c kidney areas for counting activities, and d final boundaries of the kidney

To validate the proposed approach, three medical experts manually performed the segmentation of images. Segmented data sets of experts with a Dice similarity index [28] of above 80% were used in the available data sets, as the gold standard for the rest of assessments.

In this regard, various measures of performance were utilized. The parameters and the equations are listed in Table 1 [29].

Table 1.

Various measures of performance

Measure	Definition
Sensitivity	True positive/(True positive + False negative)
Specificity	True negative/(True negative + False positive)
Precision	True positive/(True positive + False positive)
Accuracy	(True positive + True negative)/Total samples
Mean error rate	(False positive + False negative)/Total samples

The four diverse feasible results of a single prediction for a two-class case (positive or kidney tissue and or negative or non-kidney tissue) are true positive (TP_s), false positive (FP_s), false negative (FN_s), and true negative (TN_s). To this end, TP_s, FP_s, FN_s, and TN_s values are defined in confusion matrix [20] (Table 2).

Table 2.

A confusion matrix

	Actually positive (1)	Actually negative (0)
Predicted positive (1)	TPS Correctly segmented kidney tissues	FPS non-kidney tissues identified as kidney tissues
Predicted negative (0)	FNS Missed kidney tissues	TNS Correctly separated non-kidney tissues

These performance parameters for the method proposed here and thresholding with the available data sets are listed in Table 3.

Table 3.

Measures of performance for proposed and thresholding methods

Measure	Proposed method (%)	Threshold method (%)
Sensitivity	95.15	85.61
Specificity	95.33	85.23
Precision	94.91	84.33
Accuracy	94.83	86.45
Mean error rate	4.99	15.55

Another parameter to examine the proposed method is the receiver operating characteristic (ROC) curve [29]. It outperforms the traditional thresholding method in terms of accuracy (Fig. 5). The area under the ROC curve (AUC) in the proposed method is 0.974. The AUC stands for a reasonable performance statistic with the classifier systems.

Fig. 5.

The ROC curves of the proposed methods and thresholding

Additive Gaussian noise to original renal scintigraphic images was utilized to assess the performance of the proposed approach in the presence of noise. Using a constant value of mean μ and different values of variance σ² such as 0.2 and 0.4, it was found that the noisy renal data sets were segmented properly using the proposed method. A comparison between the proposed approach and thresholding method in presence of Gaussian noise with μ = 0 and σ² = 0.2 is shown in Fig. 6 and Table 4.

Fig. 6.

Initial images of renal scintigraphy data from two adult cases, from top to bottom, a original images, b noisy image, c result of kidney segmentation using the proposed method, and d traditional method

Table 4.

Performance measure of proposed and threshold methods in presence of Gaussian noise

Measure	Proposed method (%)	Thresholding method (%)
Sensitivity	90.46	78.93
Specificity	91.83	81.32
Precision	89.19	80.18
Accuracy	90.25	82.76
Mean error rate	7.21	18.75

A comparison between renogram curves which extracted based on the proposed method and a commercial one is shown in Fig. 7. As shown in the renogram curves (left), the proposed method is able to extract the same shape of the curve. The clinical renogram curve depends on initial contours around the kidneys which are selected by specialists. Therefore, human error affects the results. The proposed method performs automatically which can decrease human error for getting renogram curves, and also gives useful information to the specialists in the short time. The trend of renogram curve plays the main role of diagnosis.

Fig. 7.

Renogram curves from two patients, from left to right, a proposed method, b commercial one. Renogram curve shows activity of kidneys. The red curve shows left kidney and the green one shows right kidney. The trend generated by two methods (proposed and commercial) is identical and accordant

4. Discussion

As mentioned, the objective of the present study was performing a practicable and reliable contour for kidneys in renal scintigraphy data sets. An automatic ROI selection technique specifically designed for our renal scintigraphic application was used. In addition, an Otsu’s thresholding and integral projection to estimate the initial ROI of each kidney for decreasing processing time and error. Since renal images have different intensity ranges, acceptable results were not obtained by applying conventional techniques, such as thresholding. Moreover, choosing different threshold values was time-consuming. On the other hand, Otsu’s algorithm managed to determine the required values in an automatic manner.

As to providing an initial closed curve for the kidney, an enclosing rectangle automatically constructed ROI for each kidney. As shown in Fig. 8, the proposed techniques presented acceptable results in problematic cases. In this case, the kidney is located below its normal position (ectopic kidney). Separation refined lines are exactly Overlaying with the original images. So, left and right kidneys are segmented correctly. In addition, TAC could be accurately created in renal ptosis and patient motion during study.

Fig. 8.

A problematic renal scintigraphy image for one patient (ectopic kidney which shows right kidney is located in the abnormal place), a separation refined lines are overlaid with original image, b) left segmented kidney, and c right kidney segmented correctly

As an energy optimization technique, the variational level set is a region-based model that is developed using global image information. In addition, it is featured with many advantages compared to edge-based models and convectional level set. First of all, it does not need the image gradient, so that it demonstrates better performance with the images with weak boundaries. Moreover, the sensitivity is notable lower with the location of initiation close curves. It is also notable that it is highly robust to noise. The variational level set was efficient in generating continuous and reliable kidney contours. Figure 9 shows the results of a typical renal scintigraphic image following implementing the level set and variational level.

Fig. 9.

A typical renal scintigraphy image, from left to right, a original data, b convectional level set, and c proposed results

In renal scintigraphy procedure, nearby interfering activities such as liver and spleen can distort kidney border. The proposed method perfectly identified these and separated them from kidney. Final boundaries of renal scintigraphic images after applying thresholding and proposed method are shown in Fig. 10. Intestinal activity is diagnosed and incorrectly encompassed in ROI region of the kidney by thresholding method but, they are perfectly identified and eliminated in the process of kidney segmentation by the proposed method.

Fig. 10.

Problematic renal scintigraphy data from three adult cases, from top to bottom, a original data, b result of thresholding, and c the proposed method. Activity of adjacent organs as liver and spleen could interface with renal activity, and erroneously by thresholding method included the ROI. This problem could be accurately identified and prevented during kidney segmentation by the proposed technique. Activity of spleen is very close to renal activity, which is accurately separated by the proposed method

As indicated by the experiments, the proposed framework has a reasonable and practical robustness with the two essential problems in nuclear medicine image analysis, including noise and low-level details in automatic mode.

5. Conclusion

Segmentation of the kidneys from renal scintigraphic images is a major and changing step in order to generate the relative renal function in clinical practice. Today, in the clinical method, ROI has been selected semi-automatically by specialists. Thus, the results could be affected by human error. The proposed method performs automatically without needing to be interfered by humans. Therefore, it can reduce inter-observer and even intra-observer variability and give reliable results.

An automatic method was proposed for kidney segmentation in dynamic renal scintigraphic images using a variational level set technique. A novel initialization method was first introduced to establish the initial contour for renal scintigraphy procedure. Afterwards, it was utilized to determine the final contour by a variational level set.

The proposed method was with the manual outline of kidney boundaries done by specialists. Segmentation of the kidney using all the present data set was correct. Moreover, the method was reasonable for noisy and low-resolution images especially with the challenging cases with close interfering activities, including kidney activity. As shown by the findings, the modified variational level set algorithm had a better performance compared with the traditional thresholding and level set methods in terms of automated segmentation of kidney in renal scintigraphic images.

This article shows that the proposed method could precisely segment the renal contour in dynamic renal scintigraphy. This is especially valuable when (1) background activity is high due to renal insufficiency. (2) Liver and spleen activity interface with renal activity due to local proximity of organs. (3) Patient restlessness and motion are unavoidable during the study. (4) Technologist is unskilled for drawing the ROI.

We believe that segmentation of scintigraphic images in a completely automatic process still is an open research subject. Future work can be done on kidney segmentation in large-scale clinical data and datasets acquired by SPECT/CT. Quantitative SPECT/CT can be useful to achieve more accurate and reliable measurement of GFR [12]. Developing a computer-assisted system for analysis of scintigraphic images is another future plane.

Biographies

Arghavan Rahimi received the B.S. and M.Sc. degrees in Medical Radiation Engineering from Islamic Azad University, Science and Research Branch (SRBIAU), Tehran, Iran in 2013 and 2016, respectively. Her research interests include Medical Image Analysis.

Mohammad Hosntalab obtained his Ph.D. degrees in Medical Radiation Engineering from Islamic Azad University, Science and Research Branch (SRBIAU), Tehran, Iran. Currently, he is an Assistant Professor in the Department of Medical Radiation Engineering at SRBIAU and guiding Ph.D. students. His research interests include Medical Imaging, Medical Image Processing & Analysis, Deep Learning.

Farshid Babapour Mofrad obtained his M.Sc. and Ph.D. degrees in Medical Radiation Engineering from Islamic Azad University, Science and Research Branch (SRBIAU), Tehran, Iran. Currently, he is an Assistant Professor in the Department of Medical Radiation Engineering at SRBIAU and guiding Ph.D. students. His research interests include Medical Imaging, Image Processing & Analysis, Statistical Shape Modeling, and Machine Learning.

Mahasti Amoui obtained her Board of nuclear medicine in 1995 from Tehran Medical University, Tehran, Iran. Currently, she is an Associate Professor and Head of Nuclear Medicine Department, Shohada-e Tajrish Hospital, Shahid Beheshti University of Medical Sciences, Tehran, Iran and guiding Ph.D. students.

Ulas Bagci is an Associate Professor (with tenure) at the Northwestern University’s Radiology and Biomedical Engineering Department at Chicago. His research interests are artificial intelligence, machine learning and their applications in biomedical and clinical imaging.