Segmentation of the thoracic aorta in contrast enhanced CT images

Challenge

To develop AI algorithms for the segmentation of thoracic aorta in contrast enhanced cardiac computed tomography (CT) images.

Contrast-enhanced CT images are the preferred method for the assessment of thoracic aorta including visualization of pathologies (e.g. aortic aneurysm) and measurement of diameters at specific levels.

Accurate segmentation and visualization of aortic root is of special interest in cardiac CT imaging. Aortic root is the proximal portion of ascending aorta beginning at the level of aortic annulus and extending to sinotubular junction. Between each commissure of the aortic valve there are usually three aortic sinuses.

Automatic detection of the aortic root is an important element for further AI data processing. The difficulty in automatically segmenting the aortic root is twofold: first, the aortic valve is poorly visible and classical methods of segmentation based on thresholding usually fail. Secondly, the valve can be opened and then the contrast connects directly to the left ventricle. In both cases, radiologists and cardiologists can rely on the general anatomical features of aortic sinuses/aortic annulus to extrapolate the location of a poorly visible aortic valve. The challenge is to create an algorithm that incorporates this intuition and correctly segments the aortic root in both cases.

ML generated segmentation of thoracic aorta. Note the accurate segmentation of aortic cusps.

What we did

The Graylight Imaging team was faced with the need to develop their own algorithm. The team consisted of experts from various fields – software engineers, researchers, clinicians, physicists.

Data for the algorithm were prepared using classic image segmentation techniques such as Fast Marching Methods. The results of the prepared ground-truth were used to create an algorithm based on standard machine learning solutions.

Results

We developed an AI algorithms for the segmentation of thoracic aorta on contrast enhanced cardiac CT images. The average Sørensen-Dice coefficient of the model was above 0.98. The model is robust across images of different quality and produces meaningful results even in the case of open valve.