QCard-NM: Developing a semiautomatic segmentation method for quantitative analysis of the right ventricle in non-gated myocardial perfusion SPECT imaging

Segmentation algorithm

In our proposed approach, we named it QCard-NM, the LV is segmented first, and RV segmentation starts based on the position of the LV. Based on the RV shape in the MPI scan, an appropriate model is generated for RV segmentation, as discussed later in this section. Epicardial and endocardial surfaces are then determined via an iterative process based on the fitted model. The segmented RV volume is then used for quantification.

LV segmentation

The transversal reconstructed image is rotated based on the predefined constant angle to estimate the short-axis view. To calculate the predefined rotating angle, a group of images was automatically rotated into short-axis slices by the method described in [25]. Then, the average angle of the LV long axis of these images was considered the prior rotation angle. The image is binarized by a threshold of 50% of the voxels’ maximum count value on the upper-right side of the transversal slices. Based on the location and size of the clusters, one is selected as the LV initial cluster. The bounding box of the initial cluster is drawn and can be confirmed or changed by the user. Radial profiles (10 longitudinally and 10 latitudinally) are generated from the center of the box in 3D short-axis slices. The count profiles are averaged and its maximum is suggested as the intra-patient threshold [39]. The threshold is used to obtain the final LV cluster. An initial ellipsoidal model is fitted to the selected cluster. Radial count profiles originating from the ellipsoid center are extracted. The best point on each count profile that depicts where the epicardial and endocardial surfaces are located is chosen. A new ellipsoidal model is fitted to these selected points. These steps are iterated until an arbitrary objective is attained. The image volume is rotated based on the longest radius of the final ellipsoidal model to adjust LV orientation in the short-axis view. Finally, the valve plane position is determined based on the mid-myocardial count distribution. The flowchart of LV segmentation algorithm is provided in Additional file 1: Flowchart S.1. More comprehensive details for LV segmentation can be found in [29].

RV segmentationRV geometrical model

A semiautomatic model-based algorithm is used to segment the RV. Segmentation starts by assuming an initial spherical model centered on the septum wall. It can be proved that the spherical model is suited to the RV geometry, as follows. Figure 1 shows the short- and long-axis slices of the ventricles. “P” is denoted as the sphere center. The distance between P and the apex is called “f” which is equal to Eq. (1).

$$f = \frac\sqrt + b^ }$$

(1)

Fig. 1figure 1

Left (blue) and right (red) ventricles schema—A horizontal long-axis view—B short-axis view

RV sphericity index is defined by Eq. (2) where “a” and “b” are the short and long axis of the LV, respectively (see Fig. 1A).

If the spherical model is suitable, “f” and “b” should be equal. Based on the population reported in [40], the average end-systolic and end-diastolic sphericity indices are equal to 2 and 1.7, respectively. Consequently, according to Eq. (3), the value of \(\frac\) is 1.11 and 0.98 for the end systolic and end diastolic, respectively. “b” and “f” are close together; thus, a sphere can be an acceptable model.

$$\frac = \frac \sqrt \right)^ + 1} = \frac\sqrt + 1}$$

(3)

Another parameter for analyzing the RV geometry is the eccentricity index that is calculated as a ratio of “c”–“d” where “c” and “d” are the longest and shortest diameter of the short-axis view of the mid-part RV (see Fig. 1B). Assuming “P” as the sphere center, “m” should be close to “d” in Fig. 1B. Assume m is close to \(\frac\). Therefore, \(\frac\) is calculated by Eq. (4).

$$\frac \approx \frac} = \frac\frac = \frac }\;}$$

(4)

Again, based on the normal range of eccentricity index reported in [40], the average \(\frac\) is equal to 0.9 and 1 for the end systolic and end diastolic, respectively. It indicates that in the short-axis view, the longest and shortest radii of the mid-part RV surface model are almost identical. Thus, the spherical model is suited to non-gated MPI SPECT.

To reach the initial sphere center, the LV center point is shifted by the length of the LV ellipsoid’s smallest radius on the X-axis and one-fifth of the ellipsoid’s longest radius on the Z-axis. The preferred sphere radius is assumed to be 1.5 times greater than the smallest radius of the LV ellipsoid model. The center and radius are chosen empirically and can be modified by the user in case of model misplacement.

The LV is excluded from the short-axis image volume. Radial RV count profiles (48 longitudinally and 96 latitudinally) originating from the sphere center are extracted from the image volume. The local maxima of each profile are found. The local maximum that satisfies the following conditions is chosen: 1—Its distance to the sphere surface is less than or equal to a pixel size (6.4 mm), and 2—the voxel count value is more than one-quarter of the maximum LV count value. The corresponding voxel count value called Cprofile is used to determine the surface points. It should be mentioned that about 2000 of the total 4608 profiles satisfy these conditions. For those profiles, epicardial and endocardial surface points are chosen based on a threshold calculated by Eq. (5).

$$} = \frac - C_}}} } \right)}}C_}}}$$

(5)

where Cmax is equal to 50% of the maximum LV count value, voxels above the threshold are assumed as RV myocardium per profile, and the outer and the inner parts of the myocardium are epicardial and endocardial surfaces, respectively. Delineating epicardial and endocardial surfaces depends on the Cmax and Cprofile values differences, as well as Cprofile value, as shown in Eq. (5). Following this thresholding step, the distance between each surface point and the sphere is measured along each profile. Those endocardial and epicardial surface points whose distances are more than two pixels size (12.8 mm) are considered far-off points. The far-off points do not contribute in the proceeding steps. Along each profile, the mid-myocardial surface point is defined in the middle of the epicardial and endocardial points. The new sphere is fitted to these mid-myocardial points by the least square method. This is repeated until sphere equation coefficients reach convergence. This iterative model fitting is necessary to preserve the repeatability of the algorithm. The final model is the best sphere fitted to the mid-myocardial surface.

Epicardial and endocardial surface points

The final sphere model delineates the epicardial and endocardial surface points. Radial profiles (120 longitudinally and 320 latitudinally) originating from the sphere center are extracted from the image volume. Epicardial, endocardial, and mid-myocardial surface points are delineated as described previously. A new sphere is fitted to the mid-myocardial points. Each selected epicardial and endocardial surface point which does not meet the local maxima criterion is replaced with a point from the spherical model.

The RV valve plane is determined along the LV valve plane in the vertical long-axis view and parallel to the X-axis in the horizontal long-axis view. Voxels enclosed by the endocardial surface and the valve plane are used to determine RV cavity volume. The intersection of the RV and LV valve planes is placed on the LV epicardial surface. A fifth-degree polynomial equation (with boundary condition) is used to construct the mid-myocardial surface. Polynomial is fitted to the surface points by the Levenberg–Marquardt method [41]. RV myocardium is determined by dilating the mid-myocardial surface. Epicardial and endocardial surfaces are the outer and inner surfaces of the myocardium, respectively. The surfaces are delimited by the valve plane.

Since the normal RV wall thickness is less than one voxel size (where the system resolution is 9 mm and the voxel size is 6.4 mm) in an MPI scan, the apparent RV myocardium thickness is due to a partial volume effect. As a result, the mid-myocardium surface is the best estimator for the correct RV wall position. The final visual result of the segmented RV is the 3D mid-myocardium surface. Figure 2 shows the segmented LV and RV contours in 2D slice view for two sample patients. Figure 3 shows a sample of 3D surface rendering of a segmented cardiac volume. Epicardial and endocardial surfaces from two different view angles are illustrated in Fig. 3. Additionally, the flowchart of the RV segmentation algorithm is presented in Additional file 1: Flowchart S.2.

Fig. 2figure 2

Two sample patients’ right and left ventricular segmentation by QCard-NM. The proposed algorithm can detect the RV even in the presence of intense extracardiac activity or when the RV is partially visible, as seen in the figure

Fig. 3figure 3

A sample of the 3D surface rendering image. Left ventricular (blue) and right ventricular (red) are segmented. Epicardial and endocardial surfaces are illustrated from two different view angles (top and bottom rows)

RV polar map and quantification

The maximum count value is chosen in each myocardium sector to generate the polar map. Each sector is a part of the myocardium volume at a specific angle from the center of the RV spherical model. RV polar map generation is the same as that previously used for the LV polar maps [42]. Figure 4 shows how the RV polar map is divided into three segments. For quantitative analysis, both the maximum and the average of the counts can be computed for each segment. RV measured count is commonly normalized to the maximum LV count. Figure 5 shows two examples of quantified polar maps demonstrating the maximal/average RV/LV uptake rate in each segment.

Fig. 4figure 4

Representation of three subsegments of the RV wall in the A short-axis view and B the polar map; The RV wall is segmented into three subsegments: anterior, lateral, and inferior

Fig. 5figure 5

The stress/rest scan for two sample patients. Patient A a normal individual with no significant CAD based on CA report. Patient B a 3VD patient based on CA report. For each individual, three rows of images are represented. First row: the short-axis image slices are represented. Second row: LV is extracted automatically by the proposed segmentation algorithm to enhance the RV visualization. Third row: the RV quantified polar map by proposed method. (The numerical values represent the maximal/average RV/LV uptake ratio)

Patient population

This study enrolled four retrospective patient populations (124 patients, 50% men). All the patients underwent Tc-99 m sestamibi MPI SPECT examination. The first population includes 24 individuals who underwent two-day rest/stress SPECT/CT MPI protocols in which CT scans were acquired in both the rest and stress phases. In this paper, we call this population as SPECT/CT dataset.

The second population consists of 20 patients who underwent prone position MPI examination immediately after the supine imaging. Prone position images were acquired at the stress phase to lessen inferior wall attenuation. This dataset is called a supine/prone dataset in this literature.

Sixty patients referred for both stress/rest MPI and coronary angiography (CA) were enrolled in the third dataset, named SPECT/Angio dataset in the text. The CA was examined within 2 months after the MPI scan. Based on the CA examination and the clinical reports, patients who were right dominant and belonged to one of the following four subgroups were enrolled in this dataset:

1.

Normal patients with no significant CAD.

2.

RCA > LAD, LCX patients with proximal or mid-RCA significant stenosis with no significant lesion in the left coronary system. (Significant stenosis: > 50%)

3.

3VD Patients with severe three-vessel diseases.

4.

LAD, LCX > RCA patients with proximal or mid-LAD or LCX significant stenosis with no significant lesion in RCA. (Significant stenosis: > 50%)

Fourth dataset, called as normalcy group in this paper, consists of 20 individuals who were deemed to be at low probability of having CAD based on the following criteria:

1.

None of the following coronary risk factors hypertension, hyperlipidemia, smoking, diabetes mellitus, and chronic kidney disease (CKD).

2.

No stress-induced symptoms with successful complete stress protocol.

3.

No evidence of fixed or reversible perfusion defect on rest/stress MPI SPECT.

Patients’ clinical data, including demographic information, risk factors, MPI, and CA results, were collected from the electronic medical records and are presented in Table 1.

Table 1 Clinical characteristics, MPI data, and angiographic results of the patient populationsDigital phantoms

The non-uniform rational B-spline (NURB) extended cardiac-torso phantoms were generated in the XCAT package version 2.0 [43]. Thirty-one phantoms (16 adult males and 15 adult females) were generated with various heart sizes and abnormalities to simulate non-gated SPECT scans. (The RV cavity volume ranged from 40 to 220 ml, and defects were located on the basal and apical RV wall with three different severity levels.) The XCAT software also provided the truth values of all chambers’ volumes.

The Monte Carlo simulation program, SIMIND version 6.2.1, simulated summed myocardial perfusion images based on the XCAT phantoms and the corresponding attenuation maps [44]. Parameters of the SIMIND software were set to model the Siemens Symbia T2 hybrid SPECT/CT gamma camera (Symbia T2) (Siemens Medical Solutions Inc., Hoffman Estates, IL., USA). The Imaging protocol was set the same as the one we used in the clinic.

SPECT acquisition and processing protocol

A two-day Tc-99 m sestamibi protocol was used to perform the rest/stress scan. In both the rest and stress phases, the administered doses were based on the weight of the patients (8 MBq/kg) [45]. Rest acquisition started 60 min after the tracer was injected. Pharmacological stress was induced by an infusion of dipyridamole or dobutamine. Three minutes after the dipyridamole slow infusion, the radiotracer was injected. Stress acquisition started 45–90 min after the injection.

The SPECT studies were acquired with a dual-head detector camera, with low-energy, high-resolution collimators, a 20% symmetrical window at 140 keV, and a 64 × 64 matrix size. For each scan, 32 projections with 28 s per projection were acquired. Images were reconstructed using OSEM iterative method (four iterations and four subsets [46]), and the Butterworth post-reconstruction filter with order = 5 and cutoff = 0.5 was applied to smooth the images by QPS software v2015.1.

Images in SPECT/CT dataset were acquired by utilizing Siemens Symbia T2 dual-headed SPECT/CT gamma camera. CorCam Gamma Camera System (DDD-Diagnostic, Denmark) was used to acquire the images of patients in three other datasets. No attenuation, scatter, and detector response corrections were applied to maintain the generality of the proposed approach.

Evaluation strategies

Four strategies were performed in this paper to evaluate the performance of the proposed RV segmentation algorithm.

1.

Spatial similarity assessment An experienced nuclear medicine physician analyzed two short-axis and two horizontal long-axis views for each patient in the SPECT/CT dataset. The corresponding CT images of the selected views were also fused to the SPECT images to improve certainty. The RV contour drawn manually by the physician in each slice was compared to the contour determined by QCard-NM algorithm.

2.

Repeatability assessment The supine/prone dataset was used to assess the repeatability of the proposed algorithm and to compare it with the QPS algorithm. The QCard-NM and QPS algorithms were applied to the dataset images, and the RV cavity volumes were calculated for both the supine and prone stress images. In an ideal situation, the RV cavity volume calculated on supine and prone imaging would be the same. Considering this, the repeatability of QCard-NM and QPS was assessed in this study. Additionally, in this dataset, the RV cavity volume measured by the QCard-NM and QPS algorithms was compared to each other.

3.

Digital phantom assessment Phantom studies were utilized to quantify the accuracy of the proposed algorithm’s output. The XCAT phantoms were fed into the SIMIND software, which simulated the non-gated SPECT. The SIMIND outputs were reconstructed, and transversal images were segmented using the QCard-NM and QPS algorithms. The calculated RV cavity sizes were compared to their actual sizes (originating from the XCAT phantom).

4.

RV/LV uptake ratio assessment RV-to-LV uptake ratio (RV/LV) is defined as the RV pixel count divided by the maximum LV count. It has been demonstrated that CAD is related to the maximal RV/LV uptake ratio [5, 6, 21]. The RV/LV uptake ratio was manually generated in the previous articles from the RV lateral segment (free wall). In the fourth assessment, we investigated whether it is possible to classify patients with CAD by utilizing the proposed semiautomatic RV segmentation and quantification approach. The third and fourth datasets were used in this part. Initially, to determine the correctness of RV/LV quantitation, an expert physician visually identified those patients with high RV/LV uptake ratio. Quantitative values obtained from QCard-NM were compared to the physician’s visualized categorization. Then, a study similar to [5] was conducted to investigate the capability of semiautomatic RV quantitative analysis for classifying those patients with CAD (with no significant stenosis in RCA)

Statistical analysis

All statistical calculations were performed by use of MedCalc statistical software Version 20.104 (MedCalc, Mariakerke, Belgium). A P value less than 0.05 was considered significant.

Physicians’ manually drawn contours of the RV myocardium were compared with QCard-NM's determined contours using the Dice similarity coefficient (CSD) [47].

Mean absolute percentage error (MAPE) was utilized to show the difference between the RV volumes measured in prone and supine scans.

$$} = \frac\sum \left| }} - x_}} } \right)}}}} + x_}} } \right)}} }}} \right|$$

(6)

where \(x_}}\) and \(x_}}\) are the measured RV volumes in supine and prone scanning, respectively.

Besides, the coefficient of repeatability (CR) was calculated as 2.77 times the within-subject standard deviation. Bland–Altman analysis was utilized to study the pairwise percentage RV volume difference between prone and supine scanning [48]. Means, as well as 95% limits of agreement (LoA), were reported. Since the Shapiro–Wilk test accepted the normal distribution hypothesis of supine and prone scanning datasets, the paired sample t test was applied to compare the results.

Moreover, in this dataset, the Bland–Altman graph and paired sample t test were also used to compare the measured RV volumes between the QCard-NM and QPS algorithms.

In the phantom study, the measured RV volumes were compared to the simulated actual sizes with the MAPE metric formulated as Eq. 7:

$$} = \frac\sum \left| - x_}}} } \right)}}}}} }}} \right|$$

(7)

where \(x_ \) and \(x_ \) are the measured RV volume and the phantom actual size, respectively.

Scatter plot was presented to see the relationship between the measured and the actual volumes. Linear regression analysis was performed to extract the trend line. Pearson correlation coefficient (r) was used to explore the relationship between the measured RV volumes and the phantom actual sizes.

The Shapiro–Wilk test confirmed the SPECT/Angio dataset follows a non-normal distribution. Therefore, in the SPECT/Angio dataset, RV/LV uptake ratio was analyzed based on the visibility of the RV wall in stress MPI SPECT using the Mann–Whitney test. RV visibility was distinguished by an expert physician. Continuous data were presented as medians with interquartile range (IQR) and compared using the nonparametric Kruskal–Wallis ANOVA. The post hoc Conover test was used to examine the between-group differences. Box-and-Whisker plots were also plotted. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the diagnostic performance of RV/LV uptake ratio to distinguish those patients with 3VD or significant LAD or LCX stenosis (with no significant RCA stenosis). The area under the ROC curve (AUC), along with sensitivity, specificity, and accuracy were reported in this statistical analysis.

留言 (0)

沒有登入
gif