Echocardiography has long been an important tool for diagnosing heart defects and diseases. However, its usefulness has been hampered by difficulties in quickly and accurately identifying the myocardium (heart muscle) in echocardiograms. Unfortunately, echocardiography images are of relatively low quality, with only about 40 percent of the clinical data considered of sufficient quality for automated analysis. This problem is particularly pronounced in 3-D echocardiography, where the large number of tissues that have similar appearance to the myocardium, including adjacent muscles and bright vessel walls, complicate the discrimination even more. Finally, the sheer amount of the data contained in a 3-D echocardiography study should be processed in a matter of a few seconds—ideally in real time—to be optimally useful in clinical practice. Most, if not all, current methods fail to meet this timeliness criterion.
The University of Oxford collaborated with Microsoft Research Cambridge to investigate automated methods for segmenting 3-D echocardiography to assist cardiologists in assessing heart performance. The initial plan was to combine the graph-cuts framework developed at Microsoft Research with 3-D echocardiographic fusion, a technique for improving image quality being developed by Oxford. This plan quickly led to the idea of applying random forests to ultrasound image segmentation. (Random forests are discriminative classifiers developed recently in the machine-learning community.) The idea was to investigate whether the power of machine learning/training can lead to good segmentation results on medium-to-low quality data without the need for image fusion.
Recently, we have focused on using the temporal features, rather than just the intensity information, in a 3-D random forests framework to improve segmentation accuracy. We also are looking to establish efficient ways to train random forests for 3-D analysis. The project has demonstrated that the use of random forests allows us to obtain accurate delineations for the entire 3-D cardiac volume in a matter of seconds on a central processing unit (CPU), or even in real-time on a graphics processing unit (GPU). This class of techniques was also used by Andrew Blake and the team at Microsoft Research to produce the Kinect skeletal tracking technology. It is an example of how fundamental machine-learning research has widespread applications, ranging from medical imaging to consumer devices.
The improvement in the efficacy of echocardiography—a non-invasive diagnosis technique with no known risks or side effects—has major applications in the diagnosis of congenital heart disease and other cardiac conditions.
Learn more about this research:
- Random Forest Classification for Automatic Delineation of Myocardium in Real-Time 3D Echocardiography (PDF)