Clinically Optimized Automatic Detection of Artefactual Strain Curves

Background:

Echocardiography is the chosen imaging modality in the clinical practice. Global left ventricular function is commonly evaluated by measuring ejection fraction and regional function by the visual assessment of wall thickening. This qualitative assessment is subjective and limited to the resolution of a human eye. Strain Imaging is a novel technique based on Speckle Tracking Echocardiography (STE) that enables quantification of myocardial deformation thorough generation of ‘strain curves’.

Despite the huge number of publications concerning strain imaging, only the peak global function assessment is currently in clinical practice. A main obstacle is the uncertainty about the accuracy of the strain measurements: at present there is no ability to discern between physiological (normal or pathological) curves to artefactual curves, caused by poor tracking.

Methods:

A supervised machine learning and physiological constraint based algorithm for strain curve classification into physiological or artefactual. The data set includes healthy 415 patients, each with 3 clips of the 3 standard 2D longitudinal views.The clips were processed with an in house STE software, termed ‘K-SAD’, that has unique ability to generate independent strain curves for the 3 myocardial layers in the 6 myocardial segments. The curves were labeled by two experts separately. The proposed classification algorithm is fully automatic and treats every patient individually.

Results:

The mean patients accuracy was found to be 98% with a standard deviation range between 94.7% to 100%.

Conclusion:

An implementable and efficient algorithm for detecting artificial strain curves among healthy individuals is proposed as a necessary step for similar separation in pathological conditions.