Development and Evaluation of Two Interventricular Sulcus Extraction Methods for Cardiac PET Jizhe Wang, Tao Feng, Benjamin M.W. Tsui, Fellow IEEE Dept. of Radiology, Johns Hopkins University, Baltimore, MD, USA Abstract– The goal is to develop and evaluate two methods to extract interventricular sulcus (IS) from 4D cardiac gated (CG) myocardial perfusion (MP) PET images for use in quantitative cardiac motion estimation. In Method 1, the entire myocardium was first segmented from the 3D MP PET image of each CG frame by 3D region growing. The left ventricle (LV) was then extracted by 3D image erosion and dilation, and the right ventricle (RV) by subtracting the LV from the segmented myocardium. Finally, the IS was identified from the overlap of the extracted LV and RV. In Method 2, the inner boundary (InB) of the RV was first extracted by segmenting the blood pool (BP) within the RV using 3D region growing. The septal side of the BP boundary, or the LV outer boundary (OutB) within the RV, was separated out. We then identified segments of the LV OutB on the anterior (Ant) and posterior (Post) sides of the RV. The three extracted line segments were fitted with a B-spline curve while the lateral side of the RV InB was extrapolated using B-spline. The intersection points of the fitted curves were identified as the Ant and Post IS which, when assembled from all short-axis images, formed entire IS. The two IS extraction methods were applied to realistic CG MP PET images simulated from the 4D XCAT phantom at PET system resolution from 0.6mm to 4.5mm for 4 CG frames. The accuracy of the extracted IS were compared with the true IS from the XCAT. For both methods, the errors of the extracted IS locations increased with poorer PET system resolution with the Ant IS showing lower accuracy than the Post IS. Method 1 achieved lower accuracy than Method 2 and failed to provide reliable estimates at 4.5mm system resolution. We conclude the B-spline based interpolation and extrapolation curve fitting method was capable of extracting the IS with high reliability and accuracy from the 4D CG MP PET images obtained from state-of-the-art and PET systems and will be useful in the cardiac motion estimation.
I.
the potential benefit that the tracking of IS may bring in cardiac image analysis. Although advance in PET imaging techniques has bring the cardiac feature increasingly visible, the extraction of IS from cardiac PET images is still challenging due to the low system resolution and image noise. Moreover, there has been very few research on the extraction of this feature, and how the image quality may influence the extraction accuracy. In this project, we developed and evaluated two feature extraction method to extract the IS from 4D gated cardiac PET images for use in quantitative cardiac motion estimation. II. METHODS A. Method 1 The method was proposed in [1] as the first attempt to extract the IS for use in cardiac motion estimation. First, we used the 3D region growing method to segment out the whole myocardium from the background. Noticing that the thickness of the right ventricle is distinctively smaller than that of the left ventricle, we applied 3D erosion to the extracted myocardium followed by 3D dilation to remove the right ventricle and keep only the left ventricle. After successfully separating the right ventricle and the left ventricle, we identified the pixels that connect these two parts on each short-axis slice as the sulcus points. Fig. 1. (a) shows the extraction process on one slice of the phantom image. After extracted the interventricular sulcus points on each slice, the entire sulcus curve as in Fig. 1. (b) can be obtained by connecting these points and smoothing the curve.
INTRODUCTION
The interventricular sulcus (IS) is the line of intersection between the left ventricle (LV) and right ventricle (RV) of a human heart. As a prominent anatomical feature in human heart, it moves during the heart beat and hence provides important information about the cardiac motion, especially motion that is difficult to detect using traditional method. Methods have been proposed to use the motion information of the IS in feature-guided cardiac motion estimation [1] for improved accuracy. However, very little attention is paid on
Manuscript received December 10, 2016. Jizhe Wang, Tao Feng and Benjamin M. W. Tsui are with the Johns Hopkins University, Baltimore, MD, USA (telephone: 443-287-2425, e-mail:
[email protected]).
Fig. 1. (a) Extraction of the IS from frame 1 of the phantom image. The blue dot indicates the anterior interventricular sulcus (AIS), while the red triangle indicates anterior interventricular sulcus (PIS). (b) The extracted IS from frame 1 of the phantom image. Blue circles indicates the PIS while the red stars indicate the AIS.
B. Method 2 This method is based on the intersection of the estimated boundaries of the LV and LV using B-spline fitting. The first
step was to identify and separate the boundary of the LV and the RV from the short-axis slice images. An estimate of the inner boundary of the RV was obtained by segmenting the blood pool (BP) within the RV using the 3D region growing method. The shape of the extracted BP was concave at the septal wall side and convex at the lateral side. By subtracting the BP from its convex hull, the septal side of the BP boundary was identified as a segment of the useful outer LV boundary within the RV. We then identified the central segment of the LV boundary outside and on the anterior and posterior sides of the RV from the LV segmented from the original image using 3D region growing. Meanwhile, the segment of the BP boundary on the lateral side of the RV minus the area adjacent to the IS intersection points was identified by subtracting the central segment of the LV boundary from the whole boundary of the BP in the RV. With reduced image resolution, the sulcus point where the left and right ventricle boundary meet is blurred out and becomes less sharp and therefore unreliable. In this situation, the unreliable points near the sulcus point were removed. The three extracted outer LV segments were fitted with a B-spline curve that passed through the IS intersection points of the short-axis image slice, while the lateral boundary of BP was extrapolated using B-spline curve fitting. The two crossings of the extended BP boundary with the fitted outer LV boundary were determined as the anterior and posterior IS intersection points. This process is shown in Fig. 2.
The procedures were repeated for all short-axis image slices of the 3D cardiac PET image at each cardiac-gated frame to obtained the anterior and posterior segments of the IS over the entire heart. The final result is two sulcus curves shown in Fig. 3.
Fig 3. 3D display of extracted IS for all slices, true location of the sulcus is shown for comparison.
III. MATERIALS Four frames of gated cardiac images in voxel size of (0.6 mm)3 with activity concentration based on the FDG tracer uptake measured from the clinical PET scan of a normal patient were generated using the 4-D XCAT phantom. These frames were uniformly distributed over a one second cardiac cycle, with frame 1 as the end-diastolic (ED) phase and frame 3 as the end-systolic (ES) phase. The STIR simulation software [2] simulating a virtual highresolution PET scanner was employed to generate noise-free cardiac-gated myocardial PET projection data from each image frame. Noise-free scatter was generated using MonteCarlo simulation of clinical PET scanner and added to the high-resolution projection data after interpolation. Four system resolution levels were simulated: 0.6mm, 1.5mm, 3.0mm and 4.5mm. Lower system resolution was generated by blurring the projection of 0.6mm system resolution with Gaussian filter. The projections were reconstructed with OS-EM algorithm. IV. RESULTS AND DISCUSSIONS
Fig. 2. (a) A short-axis slice of the XCAT phantom with activity in myocardium and liver. (b) Blood pool inside the right ventricle segmented using 3D region growing method. (c) Residual shape after subtracting the blood pool from its convex hull, from which the septal boundary is identified. (d) Separated septal boundary and lateral boundary shown in pink and blue respectively. (e) Septal boundary with added points from left ventricle boundary and lateral boundary shown in pink and blue respectively. (f) B-spline curve fitting of septal boundary and extrapolation of lateral boundary on one shortaxis slice. The intersection is marked by the yellow points.
For quantitative evaluation of the IS extraction results, we define the error of estimated sulcus location as 1
𝐸 = ∑𝑁 𝑖=1‖𝑳𝑒𝑠𝑡 (𝑖) − 𝑳𝑡𝑟 (𝑖)‖2 , 𝑁
where 𝑖 is the short-axis slice number, N is the total number of short-axis slices, 𝑳𝑒𝑠𝑡 (𝑖) and 𝑳𝑡𝑟 (𝑖) are the estimated 2D location of sulcus point and the true location of sulcus point on short-axis slice i, respectively; ‖ ‖2 is the l2 norm. The true sulcus point’s location was provided by the XCAT phantom. The results for both methods in noise-free simulation data were summarized in Table I. The errors for AIS and PIS were
calculated separately. With both feature extraction methods, the error in IS extraction increases as the system resolution degrades. Under the four simulated system resolutions, Method 2 generally has higher accuracy than Method 1. Especially, for the 4.5mm resolution cardiac images, Method 1 was unable to extract the sulcus location. Another distinctive phenomenon is that the accuracy of PIS is almost always higher than that of AIS under four system resolutions and in four cardiac frames. The degradation of its accuracy under poor image resolution is much less severe than that of the AIS. This is caused by the difference in the shape of the two corners of the BP in the RV. The anterior corner is an acute angle, making it easier to be blurred out under poor system resolution. The posterior corner is an obtuse angle, less susceptible to resolution degradation. Table I. IS extraction error (mm) using two IS extraction methods for simulated cardiac PET images of four system resolutions
Frame 1 Frame 2 Frame 3 Frame 4
Frame 1 Frame 2 Frame 3 Frame 4
Frame 1 Frame 2 Frame 3 Frame 4
Frame 1 Frame 2 Frame 3 Frame 4
Method 1 Method 2 0.6 simulation noise free AIS PIS AIS PIS 0.91 0.91 1.04 0.85 1.28 1.59 1.06 0.49 0.86 0.70 1.13 0.49 0.84 0.63 1.02 0.49 1.5 simulation noise free AIS PIS AIS PIS 1.60 1.64 1.52 0.83 1.49 1.02 1.78 0.66 1.54 1.66 1.81 0.72 1.66 1.53 1.80 0.69 3 simulation noise free AIS PIS AIS PIS 4.69 2.55 2.39 1.01 2.45 1.97 2.90 0.96 2.61 3.18 3.13 1.12 3.77 2.65 2.54 0.85 4.5 simulation noise free AIS PIS AIS PIS fail fail 3.53 1.02 fail fail 4.27 1.39 fail fail 4.30 1.64 fail fail 3.70 1.10
Accurate extraction of the IS is crucial for detecting the circumferential motion of the heart. To understand how the feature extraction accuracy can influence motion analysis of the sulcus, we calculated the motion of the posterior sulcus between each two frames extracted with method 2 based on the linear assumption of the longitudinal motion, which states that the longitudinal motion increases linearly from the apex to the base of the heart with 0 value at the apex. The two basal end-points of the sulcus in each two frames were matched, and then other points were matched proportionally. Three cardiac motion components – radial, circumferential and longitudinal
motion – were calculated based on the point-to-point correspondence and the shape of the heart. Fig. 4 to Fig. 6 show the three motion components from frame 1 to frame 2 of the PIS calculated from phantom images. On the curves, each data point represents an extracted sulcus point. The long-axis slice number from small value to large value represents sulcus point location along cardiac long-axis from apex to base. Notice that the circumferential motion is measured in degree, while the radial and longitudinal motion are measured in pixel. Here we show the radial and longitudinal motion in pixels instead of mile-meter, because feature extraction is carried out in pixelated cardiac images with 0.6mm pixel size. degree 8 6 4 2 0 -2 -4 -6 50
Vc_est
Vc_Truth
long-axis slice 100
150
Fig. 4 Circumferential motion of PIS from frame 1 to frame 2 calculated from phantom images. The true circumferential motion is shown in darker green as reference. pixel 7 6 5 4 3 2 1 0 50
Vr_Est
Vr_Truth
long-axis slice 100
150
Fig 5. Radial motion of PIS from frame 1 to frame 2 calculated from phantom images. The true radial motion is shown in darker red as reference. pixel 0
Vl_Est
Vl_Truth
-5 -10 long-axis slice -15
50
100
150
Fig 6. Longitudinal motion of PIS from frame 1 to frame 2 calculated from phantom images. The true longitudinal motion is shown in darker blue as reference.
The three motion components of the PIS calculated from the phantom images are reasonably accurate. These three figures reveal the cardiac motion pattern of human heart. Especially, Fig. 5 indicates that the heart is twisting in the opposite directions along the long-axis. This motion component is difficult to observe from cardiac PET images. Although the motion of PIS extracted from phantom images is reasonably accurate, it is not the case in simulated cardiac PET images, especially when the system resolution gets worse. The following figures show the circumferential motion of PIS from frame 1 to frame 2 extracted from noise-free simulation images of different system resolution. degree 8 6 4 2 0 -2 -4 -6 65
Vc_Est
Vc_Truth
long-axis slice 115
165
Fig.7 Circumferential motion of PIS from gated Frame #1 to #2 extracted from noise-free simulation data of 0.6mm system resolution. The true circumferential motion is shown in darker green. degree 8 6 4 2 0 -2 -4 -6 65
Vc_Est
Vc_Truth
long-axis slice 115
165
Fig.8 Circumferential motion of PIS from gated Frame #1 to #2 extracted from noise-free simulation data of 1.5mm system resolution. The true circumferential motion is shown in darker green. degree 8
Vc_Est
Vc_Truth
6 4 2 0 -2 -4
long-axis slice
-6 65
115
Fig.9 Circumferential motion of PIS from gated Frame #1 to #2 extracted from noise-free simulation data of 3.0mm system resolution. The true circumferential motion is shown in darker green. degree 8
Vc_Est
6 4 2 0 -2 -4
long-axis slice
-6 65
115
165
Fig.10 Circumferential motion of PIS from gated Frame #1 to #2 extracted from noise-free simulation data of 4.5mm system resolution. The true circumferential motion is shown in darker green.
For circumferential motion, there is a visible trend from high resolution to low resolution. When the system resolution is very high and same as the pixel size of the phantom image, the circumferential motion is close to the truth. However, as system resolution becomes lower, the slope of the circumferential motion decreases. At 4.5 mm resolution, the linear regression line of circumferential motion is almost flat. In other words, the “twisting” of the heart becomes less and less distinguishable as the resolution degrades. Besides the changes in motion estimation accuracy, the curves for the PIS motion also become shorter at the apical end from results of 0.6mm resolution to that of 4.5mm resolution. It gets more difficult to extract the sulcus points near the apex with lower resolution, because the right ventricle is much smaller at the apical region and hence is more susceptible to the blurring of lower system resolution. V. CONCLUSION Two cardiac feature extraction methods were proposed for extraction of the IS of the heart. Method 1 relies on separating LV and RV with image erosion and dilation, while Method 2 is based on B-spline curve fitting, interpolation and extrapolation techniques of identifiable left and right ventricle boundary information. Method 2 provides improved extraction accuracy of the IS than Method 1 for simulation data of four system resolutions, and allow its use to track its motion for aid in cardiac motion estimate from 4D gated cardiac PET images with continuing improving resolution. The results from simulation data of different resolutions show that with the degradation of resolution in cardiac PET images, the extracted motion estimate of the sulcus becomes less accurate. Especially, the circumferential motion becomes less prominent as resolution becomes lower. REFERENCE
165
Vc_Truth
[1] Jizhe, W., et al. An interventricular sulcus guided cardiac motion estimation method. in 2013 IEEE Nuclear Science Symposium and Medical Imaging Conference (2013 NSS/MIC). 2013. [2] Thielemans, K., et al., STIR: software for tomographic image reconstruction release 2. Physics in Medicine and Biology, 2012. 57(4): p. 867-883.
