Development of a semi-automated segmentation tool for high frequency ultrasound image analysis of mouse echocardiograms

Echocardiography is a widely used and clinically translatable imaging modality for the evaluation of cardiac structure and function in preclinical drug discovery and development. Echocardiograms are among the first in vivo diagnostic tools utilized to evaluate the heart due to its relatively low cost, high throughput acquisition, and non-invasive nature; however lengthy manual image analysis, intra- and inter-operator variability, and subjective image analysis presents a challenge for reproducible data generation in preclinical research. To combat the image-processing bottleneck and address both variability and reproducibly challenges, we developed a semi-automated analysis algorithm workflow to analyze long- and short-axis murine left ventricle (LV) ultrasound images. The long-axis B-mode algorithm executes a script protocol that is trained using a reference library of 322 manually segmented LV ultrasound images. The short-axis script was engineered to analyze M-mode ultrasound images in a semi-automated fashion using a pixel intensity evaluation approach, allowing analysts to place two seed-points to triangulate the local maxima of LV wall boundary annotations. Blinded operator evaluation of the semi-automated analysis tool was performed and compared to the current manual segmentation methodology for testing inter- and intra-operator reproducibility at baseline and after a pharmacologic challenge. Comparisons between manual and semi-automatic derivation of LV ejection fraction resulted in a relative difference of 1% for long-axis (B-mode) images and 2.7% for short-axis (M-mode) images. Our semi-automatic workflow approach reduces image analysis time and subjective bias, as well as decreases inter- and intra-operator variability, thereby enhancing throughput and improving data quality for pre-clinical in vivo studies that incorporate cardiac structure and function endpoints.