Tracking the movement of tissue boundaries in the heart helps cardiologists to identify abnormalities such as poor muscle contraction or reduced motion during the cardiac cycle. Currently this is a time consuming process, as cardiologists are required to manually trace the boundaries in each frame of an ultrasound video. This technology is a fully automated method for identifying these boundaries, which is based on an autocorrelation algorithm. Different structures of the heart reflect echo signals at different amplitudes, and an edge detection technique is applied to this amplitude difference thus identifying the boundaries. To increase accuracy, a machine learning technique can be used to find the optimal threshold value for identifying echo amplitude differences, even varying this threshold from frame to frame if necessary.
The time required to analyze echocardiogram data is greatly reduced if automated boundary detection can be implemented. Previous attempts in ultrasound videos have suffered from the high levels of noise present in such videos. This technology uses grayscale pixel values from adjacent frames to calculate their autocorrelation, while noise is reduced by implementing a perceptron machine learning algorithm to the boundary detection. As opposed to a manual approach, where increasing the frame rate will result in longer and more cumbersome analyses times, higher frame rates are expected to help refine the results in this automated approach. Various types of morphological and filtering operations further improve accuracy, and additionally the image is resized if pixels move out of the frame.
This technology has been tested by calculating the area of the endocardium of a left ventricle, and the relative error when compared to manual tracings is less than 1% in some cases.
Patent Pending (US 20070276245)
Tech Ventures Reference: IR 1699