The analysis of cine MR images plays an important role in the assessment of cardiac function. Quantitative analysis of cardiac function requires manual tracing of the ventricular contours by an experienced post-processor, which is time-consuming. To reduce the burden on post-processors, semi-automated and automatic algorithms have been developed. Most algorithms have been designed for the segmentation of the left ventricle (LV); however, functional parameters for the right ventricle (RV) are essential for main disease processes such as pulmonary hypertension and congenital heart disease¹.

Most segmentation algorithms applied to the heart fall into two categories: those using training data and those using low-level features, such as intensity and gradients. The main drawback of the algorithms that rely on training data is that the performance degrades with pathological cases². The algorithms that focus on low-level features may converge to an incorrect contour; for example, many experts draw the LV endocardial border around the blood pool, outside the strongest edge³. In this abstract we propose a novel segmentation algorithm for the RV in 4-chamber (4-CH) cardiac images. The new algorithm uses low-level features and incorporates user input to identify some landmarks in the RV in 4-CH cardiac images.

Given a representative cine 4-CH image of the heart, the RV is outlined as follows. First, the image is automatically cropped to include only the chambers of the heart (ventricles and atria). This is accomplished by looking at the variance of the pixels through the cardiac cycle (e.g., across 25 cardiac frames), thresholding the variance image and finding the bounding box of the largest connected component in the variance image. Once the chambers of the heart have been cropped, the post-processor is asked to identify the end-diastolic frame. This is the frame where the RV has the largest area. In the end-diastolic frame, the post-processor has to select four landmarks associated with the RV: the apex, the lateral and medial sides of the tricuspid valve (TV), and the middle of the blood pool. These landmarks are shown in Fig. 1a. Then the post-processor is asked to identify the end-systolic frame. This is the frame where the RV area is the smallest and can be identified because the TV is closed. In the end-systolic frame, the post-processor identifies the TV landmarks (lateral and medial). These landmarks are shown in Fig. 1b. Given the landmarks, the algorithm extrapolates the location of the landmarks throughout the cardiac cycle, as shown in Fig. 2, and performs a 2D segmentation on each frame of the cardiac cycle using a variation of a polar dynamic programming⁴ technique developed by our group. The final step is to smooth the 2D segmentation to ensure temporal smoothness, as shown in Fig. 3.

To evaluate the segmentation performance, we analyzed 175 images coming from 7 different patients (25 frames per patient). Images were acquired using a cine steady-state free-precession pulse sequence (TR = 2.7, 2.8 ms, TE = 1.16, 1.17 ms, acq. matrix = 168 x 192, FOV = 29.8 x 34 cm, slice thickness = 6 mm) in a Siemens 1.5T Aera scanner.

The segmentation algorithm described above was used to outline the RV in the 4-CH cardiac images. The semi-automated segmentation result was compared to the manual tracing of two experienced cardiovascular post-processors or tracers, using the mean and standard deviation of the Dice metric, which is a measure of mutual overlap between ground truth and the segmented region.

The segmentation results for three different patients are shown in Fig. 4. The corresponding Dice metric values (calculated using tracer #1 as the ground truth) are 0.913, 0.887 and 0.905. Fig. 5 shows the Dice metric matrix. The matrix shows the intertracer variability (comparison between tracer #1 and #2) as well as the Dice metric between the semi-automated algorithm and each of the manual tracers. Note that the semi-automated technique is within human variability. Specifically, the mean Dice metric between tracer #1 and the semi-automated technique is larger (i.e. higher overlap) than between tracer #1 and tracer #2. Our technique took on average 2 minutes per patient, as compared to 20 minutes per patient for the tracers.We have developed a semi-automated algorithm that accurately segments RV 4-CH MR images within human variability. The RV segmentation in the 4–CH image can be used to measure the ventricular wall motion in the RV similar to the work done by Nagel et al. for the LV⁵.