Landmark Detection

Landmark Detection in Cardiovascular Images Using Neural Networks

Dutch Technology Foundation STW: LGN 4503)
Elco Oost, MSc

Background

To determine Left Ventricular (LV) function, many cardiac patients have to undergo a catheterization procedure, in which the left ventricle (LV) is filled with an X-ray opaque contrast dye. Subsequently the beating left ventricle can be visualized by X-ray radiation in a series of about 150 to 200 images. To quantitatively measure cardiac function, currently the LV contours are drawn manually by expert cardiologists. This process is difficult, time-consuming and prone to inter- and intra-observer variability. Therefore, automatic contour detection in Left Ventricle (LV) X-ray angiograms is desired to quantitatively evaluate cardiac function.
Automatic contour detection in LV angiograms has shown to be a very difficult problem, mainly because of the large variations in clinical image quality, and the many artifacts that may occur. In order to robustly detect the left-ventricular boundaries, a-priori knowledge about the LV shape and image appearance is required in the form of an anatomical model. This project explores various novel ways to represent and use such a-priori knowledge for achieve a robust and reliable contour detection.

Goals

To develop robust, near-fully automatic contour detection for evaluation of LV function from X-ray angiograms,

To develop adaptive, knowledge driven segmentation methods.

Approach

Initially, a number of segmentation approaches were developed to automatically localize the apex and valve landmarks of the left ventricle, which were mainly based on neural-network techniques. These were selected for their adaptivity and ability to model knowledge about shape and appearance in an implicit manner. However, during the project the focus was shifted towards Active Appearance Models (AAMs), and currently, we are developing localized and multi-view Active Appearance Models to simultaneously segment end-diastolic and end-systolic views. The application of multi-view AAM's combined with a directed dynamic programming yielded robust detection, without the need for much user interaction and contour correction.

AngiotestExample4EDES

EDvar_strip_123

Click on the images to activate the movies.

Status

The project ended in March 2005. The developed algorithms have been transferred to Medis medical imaging systems, who have integrated the detection methods into their QAngio XA software.

Contact

For further information, please contact:
B.P.F. Lelieveldt PhD,
Division of Image Processing
Department of Radiology, 1-C2S
Leiden University Medical Center
P.O. Box 9600
2300 RC Leiden
The Netherlands
Tel. +31 (0)71 526 2285
Fax. +31 (0)71 526 6801
e-mail: B.P.F.Lelieveldt@lumc.nl