|Ph.D Student||Khamis Hanan|
|Subject||New Constraint-Based Approaches for Enhancing the Accuracy|
and Reliability of 2D Echocardiographic Assessment
of Left Ventricular Function
|Department||Department of Biomedical Engineering||Supervisors||PROFESSOR EMERITUS Dan Adam|
|DR. Zvi Friedman|
Cardiovascular diseases (CAD) are the cause of more than 31 percent of all deaths worldwide. The impairment of left ventricle (LV) diastolic and systolic function appears very early in the course of CAD diseases. Early and precise detection of LV myocardial malfunction may prevent function deterioration and allow prompt and appropriate treatment. Accordingly, recent CAD guidelines accentuate the need for non-invasive detection of subclinical LV dysfunction using conventional echocardiography combined with global strain measurement.
The most commonly used technique for the assessment of LV strains is two-dimensional speckle tracking echocardiography (2D STE). Recent STE algorithms provide transmural, segmental and regional myocardial mechanical measurements (e.g. strain) in addition to global values. However, since echocardiographic clips are corrupted by speckle decorrelation noise, resulting in irregular, non-physiological tissue displacement fields, smoothing and regularization are performed on the displacement data, affecting the strain results. Thus, strain results may depend on the specific implementations of 2D STE, and the software products offered by the various vendors produce significantly different results for the same clip/patient. In addition, these products do not provide a reliability tool that indicates whether the functional measurements are physiological or not to help avoid erroneous diagnosis. Finally, yet importantly, these products do not allow a fully automatic processing, since it requires manual selection of the clip of interest for analysis. These aforementioned drawbacks limit the clinical acceptance of 2D STE.
In this work, a cascade of algorithms is proposed. The first algorithm allows automatic classification of apical LV echocardiograms into the different anatomical views using a multi-stage classification algorithm, which employs spatio-temporal feature extraction and supervised dictionary learning approaches. The classification results demonstrated a promising approach despite the inter-view similarity between the classes and the intra-view variability among clips belonging to the same class.
The second algorithm introduces novel approach for estimating the mechanical changes of the myocardium along the cardiac cycle. It integrates the physiological constraint of smoothness of the tissue`s displacement field into an optimization process. The strain measurements that were produced when analyzing simulated echocardiograms and a large cohort of clinical data set indicated that the sensitivity of strain values to speckle noise, caused by the post block-matching weighted smoothing, are significantly reduced and the accuracy is concurrently enhanced.
Finally, a supervised machine learning approach is proposed as a third algorithm for the classification of the spatio-temporal strain measurements into artefactual or physiological patterns, to allow reliable determination prior to clinical diagnosis. This classification tool demonstrated a high potential of use for enhancing the reliability of the estimation of the global, transmural and regional strain measurements. It can also be implemented in any strain estimation algorithm to further enhance the accuracy and robustness of the measurements.
The proposed new set of tools may enhance echocardiographic diagnostic accuracy and allow its integration into the clinical routine.