Automatic tricuspid valve annulus extraction and measurement from computed tomography images

Abstract

Background and objective: Tricuspid regurgitation (TR) is one of the most common forms of valvular heart diseases. The morphological information of the tricuspid valve annulus (TVA) is critical in treatment planning for TR. It is necessary to extract the TVA from medical images to obtain that information, however this task is difficult and time-consuming to perform manually. In this paper, we propose a method to automatically extract and measure the TVA from computed tomography (CT) images. Methods: Our proposed method coarsely crops CT images to the region surrounding the tricuspid valve based on the right atrium and the right ventricle regions. The cropped CT images are input to a stacked hourglass network with loss function integrating the mean squared error loss, the focal loss and the shape-aware weighted Hausdorff distance loss to extract 36 landmarks on the TVA. The extraction accuracy of TVA landmarks was evaluated by five-fold cross validation using 120 CT images with manually annotated TVA landmarks. In addition, measurements of TVA morphology based on automatically extracted TVA and those based on manually annotated TVA were calculated and compared using the same measurement algorithm which provides a means to automatically generate seven measurements based on TVA landmarks. Results: Our proposed method extracted TVA inside the right heart in all CT images without any processing interruption. The mean processing time was 27.09 ± 8.65 s, and the Chamfer distance and Hausdorff distance were 2.07 ± 0.53 and 4.09 ± 1.29, respectively. The mean absolute error between the measurements based on automatically extracted TVA and those based on manually annotated TVA was less than 4 mm, which is less than the typical device size interval for surgical prosthetic valve rings in current use, for measurement items related to distance. For all seven measurement items, significant correlations (r = 0.51–0.99, p < 0.0071) were shown between the measurements based on automatically extracted TVA and those based on manually annotated TVA. Conclusions: Our proposed method was able to automatically extract and measure the TVA. This method is expected to reduce the time and effort required by physicians in treatment planning for TR.

Keywords

• Tricuspid valve
• Segmentation
• Deep learning
• Computed tomography